mira/bertalign
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Baseline alignment systems

Browse Source
This commit is contained in:
nlpfun
2021-11-28 13:59:28 +08:00
parent e033edad52
commit cc1ca021e8
34 changed files with 453434 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

5
ext-lib/bleualign/.gitignore vendored Normal file
View File

@@ -0,0 +1,5 @@
__pycache__/
*.pyc
/dist
/build
/MANIFEST

339
ext-lib/bleualign/LICENSE Normal file
View File

@@ -0,0 +1,339 @@
GNU GENERAL PUBLIC LICENSE
Version 2, June 1991
Copyright (C) 1989, 1991 Free Software Foundation, Inc.,
51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
Everyone is permitted to copy and distribute verbatim copies
of this license document, but changing it is not allowed.
Preamble
The licenses for most software are designed to take away your
freedom to share and change it. By contrast, the GNU General Public
License is intended to guarantee your freedom to share and change free
software--to make sure the software is free for all its users. This
General Public License applies to most of the Free Software
Foundation's software and to any other program whose authors commit to
using it. (Some other Free Software Foundation software is covered by
the GNU Lesser General Public License instead.) You can apply it to
your programs, too.
When we speak of free software, we are referring to freedom, not
price. Our General Public Licenses are designed to make sure that you
have the freedom to distribute copies of free software (and charge for
this service if you wish), that you receive source code or can get it
if you want it, that you can change the software or use pieces of it
in new free programs; and that you know you can do these things.
To protect your rights, we need to make restrictions that forbid
anyone to deny you these rights or to ask you to surrender the rights.
These restrictions translate to certain responsibilities for you if you
distribute copies of the software, or if you modify it.
For example, if you distribute copies of such a program, whether
gratis or for a fee, you must give the recipients all the rights that
you have. You must make sure that they, too, receive or can get the
source code. And you must show them these terms so they know their
rights.
We protect your rights with two steps: (1) copyright the software, and
(2) offer you this license which gives you legal permission to copy,
distribute and/or modify the software.
Also, for each author's protection and ours, we want to make certain
that everyone understands that there is no warranty for this free
software. If the software is modified by someone else and passed on, we
want its recipients to know that what they have is not the original, so
that any problems introduced by others will not reflect on the original
authors' reputations.
Finally, any free program is threatened constantly by software
patents. We wish to avoid the danger that redistributors of a free
program will individually obtain patent licenses, in effect making the
program proprietary. To prevent this, we have made it clear that any
patent must be licensed for everyone's free use or not licensed at all.
The precise terms and conditions for copying, distribution and
modification follow.
GNU GENERAL PUBLIC LICENSE
TERMS AND CONDITIONS FOR COPYING, DISTRIBUTION AND MODIFICATION
0. This License applies to any program or other work which contains
a notice placed by the copyright holder saying it may be distributed
under the terms of this General Public License. The "Program", below,
refers to any such program or work, and a "work based on the Program"
means either the Program or any derivative work under copyright law:
that is to say, a work containing the Program or a portion of it,
either verbatim or with modifications and/or translated into another
language. (Hereinafter, translation is included without limitation in
the term "modification".) Each licensee is addressed as "you".
Activities other than copying, distribution and modification are not
covered by this License; they are outside its scope. The act of
running the Program is not restricted, and the output from the Program
is covered only if its contents constitute a work based on the
Program (independent of having been made by running the Program).
Whether that is true depends on what the Program does.
1. You may copy and distribute verbatim copies of the Program's
source code as you receive it, in any medium, provided that you
conspicuously and appropriately publish on each copy an appropriate
copyright notice and disclaimer of warranty; keep intact all the
notices that refer to this License and to the absence of any warranty;
and give any other recipients of the Program a copy of this License
along with the Program.
You may charge a fee for the physical act of transferring a copy, and
you may at your option offer warranty protection in exchange for a fee.
2. You may modify your copy or copies of the Program or any portion
of it, thus forming a work based on the Program, and copy and
distribute such modifications or work under the terms of Section 1
above, provided that you also meet all of these conditions:
a) You must cause the modified files to carry prominent notices
stating that you changed the files and the date of any change.
b) You must cause any work that you distribute or publish, that in
whole or in part contains or is derived from the Program or any
part thereof, to be licensed as a whole at no charge to all third
parties under the terms of this License.
c) If the modified program normally reads commands interactively
when run, you must cause it, when started running for such
interactive use in the most ordinary way, to print or display an
announcement including an appropriate copyright notice and a
notice that there is no warranty (or else, saying that you provide
a warranty) and that users may redistribute the program under
these conditions, and telling the user how to view a copy of this
License. (Exception: if the Program itself is interactive but
does not normally print such an announcement, your work based on
the Program is not required to print an announcement.)
These requirements apply to the modified work as a whole. If
identifiable sections of that work are not derived from the Program,
and can be reasonably considered independent and separate works in
themselves, then this License, and its terms, do not apply to those
sections when you distribute them as separate works. But when you
distribute the same sections as part of a whole which is a work based
on the Program, the distribution of the whole must be on the terms of
this License, whose permissions for other licensees extend to the
entire whole, and thus to each and every part regardless of who wrote it.
Thus, it is not the intent of this section to claim rights or contest
your rights to work written entirely by you; rather, the intent is to
exercise the right to control the distribution of derivative or
collective works based on the Program.
In addition, mere aggregation of another work not based on the Program
with the Program (or with a work based on the Program) on a volume of
a storage or distribution medium does not bring the other work under
the scope of this License.
3. You may copy and distribute the Program (or a work based on it,
under Section 2) in object code or executable form under the terms of
Sections 1 and 2 above provided that you also do one of the following:
a) Accompany it with the complete corresponding machine-readable
source code, which must be distributed under the terms of Sections
1 and 2 above on a medium customarily used for software interchange; or,
b) Accompany it with a written offer, valid for at least three
years, to give any third party, for a charge no more than your
cost of physically performing source distribution, a complete
machine-readable copy of the corresponding source code, to be
distributed under the terms of Sections 1 and 2 above on a medium
customarily used for software interchange; or,
c) Accompany it with the information you received as to the offer
to distribute corresponding source code. (This alternative is
allowed only for noncommercial distribution and only if you
received the program in object code or executable form with such
an offer, in accord with Subsection b above.)
The source code for a work means the preferred form of the work for
making modifications to it. For an executable work, complete source
code means all the source code for all modules it contains, plus any
associated interface definition files, plus the scripts used to
control compilation and installation of the executable. However, as a
special exception, the source code distributed need not include
anything that is normally distributed (in either source or binary
form) with the major components (compiler, kernel, and so on) of the
operating system on which the executable runs, unless that component
itself accompanies the executable.
If distribution of executable or object code is made by offering
access to copy from a designated place, then offering equivalent
access to copy the source code from the same place counts as
distribution of the source code, even though third parties are not
compelled to copy the source along with the object code.
4. You may not copy, modify, sublicense, or distribute the Program
except as expressly provided under this License. Any attempt
otherwise to copy, modify, sublicense or distribute the Program is
void, and will automatically terminate your rights under this License.
However, parties who have received copies, or rights, from you under
this License will not have their licenses terminated so long as such
parties remain in full compliance.
5. You are not required to accept this License, since you have not
signed it. However, nothing else grants you permission to modify or
distribute the Program or its derivative works. These actions are
prohibited by law if you do not accept this License. Therefore, by
modifying or distributing the Program (or any work based on the
Program), you indicate your acceptance of this License to do so, and
all its terms and conditions for copying, distributing or modifying
the Program or works based on it.
6. Each time you redistribute the Program (or any work based on the
Program), the recipient automatically receives a license from the
original licensor to copy, distribute or modify the Program subject to
these terms and conditions. You may not impose any further
restrictions on the recipients' exercise of the rights granted herein.
You are not responsible for enforcing compliance by third parties to
this License.
7. If, as a consequence of a court judgment or allegation of patent
infringement or for any other reason (not limited to patent issues),
conditions are imposed on you (whether by court order, agreement or
otherwise) that contradict the conditions of this License, they do not
excuse you from the conditions of this License. If you cannot
distribute so as to satisfy simultaneously your obligations under this
License and any other pertinent obligations, then as a consequence you
may not distribute the Program at all. For example, if a patent
license would not permit royalty-free redistribution of the Program by
all those who receive copies directly or indirectly through you, then
the only way you could satisfy both it and this License would be to
refrain entirely from distribution of the Program.
If any portion of this section is held invalid or unenforceable under
any particular circumstance, the balance of the section is intended to
apply and the section as a whole is intended to apply in other
circumstances.
It is not the purpose of this section to induce you to infringe any
patents or other property right claims or to contest validity of any
such claims; this section has the sole purpose of protecting the
integrity of the free software distribution system, which is
implemented by public license practices. Many people have made
generous contributions to the wide range of software distributed
through that system in reliance on consistent application of that
system; it is up to the author/donor to decide if he or she is willing
to distribute software through any other system and a licensee cannot
impose that choice.
This section is intended to make thoroughly clear what is believed to
be a consequence of the rest of this License.
8. If the distribution and/or use of the Program is restricted in
certain countries either by patents or by copyrighted interfaces, the
original copyright holder who places the Program under this License
may add an explicit geographical distribution limitation excluding
those countries, so that distribution is permitted only in or among
countries not thus excluded. In such case, this License incorporates
the limitation as if written in the body of this License.
9. The Free Software Foundation may publish revised and/or new versions
of the General Public License from time to time. Such new versions will
be similar in spirit to the present version, but may differ in detail to
address new problems or concerns.
Each version is given a distinguishing version number. If the Program
specifies a version number of this License which applies to it and "any
later version", you have the option of following the terms and conditions
either of that version or of any later version published by the Free
Software Foundation. If the Program does not specify a version number of
this License, you may choose any version ever published by the Free Software
Foundation.
10. If you wish to incorporate parts of the Program into other free
programs whose distribution conditions are different, write to the author
to ask for permission. For software which is copyrighted by the Free
Software Foundation, write to the Free Software Foundation; we sometimes
make exceptions for this. Our decision will be guided by the two goals
of preserving the free status of all derivatives of our free software and
of promoting the sharing and reuse of software generally.
NO WARRANTY
11. BECAUSE THE PROGRAM IS LICENSED FREE OF CHARGE, THERE IS NO WARRANTY
FOR THE PROGRAM, TO THE EXTENT PERMITTED BY APPLICABLE LAW. EXCEPT WHEN
OTHERWISE STATED IN WRITING THE COPYRIGHT HOLDERS AND/OR OTHER PARTIES
PROVIDE THE PROGRAM "AS IS" WITHOUT WARRANTY OF ANY KIND, EITHER EXPRESSED
OR IMPLIED, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF
MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE. THE ENTIRE RISK AS
TO THE QUALITY AND PERFORMANCE OF THE PROGRAM IS WITH YOU. SHOULD THE
PROGRAM PROVE DEFECTIVE, YOU ASSUME THE COST OF ALL NECESSARY SERVICING,
REPAIR OR CORRECTION.
12. IN NO EVENT UNLESS REQUIRED BY APPLICABLE LAW OR AGREED TO IN WRITING
WILL ANY COPYRIGHT HOLDER, OR ANY OTHER PARTY WHO MAY MODIFY AND/OR
REDISTRIBUTE THE PROGRAM AS PERMITTED ABOVE, BE LIABLE TO YOU FOR DAMAGES,
INCLUDING ANY GENERAL, SPECIAL, INCIDENTAL OR CONSEQUENTIAL DAMAGES ARISING
OUT OF THE USE OR INABILITY TO USE THE PROGRAM (INCLUDING BUT NOT LIMITED
TO LOSS OF DATA OR DATA BEING RENDERED INACCURATE OR LOSSES SUSTAINED BY
YOU OR THIRD PARTIES OR A FAILURE OF THE PROGRAM TO OPERATE WITH ANY OTHER
PROGRAMS), EVEN IF SUCH HOLDER OR OTHER PARTY HAS BEEN ADVISED OF THE
POSSIBILITY OF SUCH DAMAGES.
END OF TERMS AND CONDITIONS
How to Apply These Terms to Your New Programs
If you develop a new program, and you want it to be of the greatest
possible use to the public, the best way to achieve this is to make it
free software which everyone can redistribute and change under these terms.
To do so, attach the following notices to the program. It is safest
to attach them to the start of each source file to most effectively
convey the exclusion of warranty; and each file should have at least
the "copyright" line and a pointer to where the full notice is found.
<one line to give the program's name and a brief idea of what it does.>
Copyright (C) <year> <name of author>
This program is free software; you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation; either version 2 of the License, or
(at your option) any later version.
This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License along
with this program; if not, write to the Free Software Foundation, Inc.,
51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
Also add information on how to contact you by electronic and paper mail.
If the program is interactive, make it output a short notice like this
when it starts in an interactive mode:
Gnomovision version 69, Copyright (C) year name of author
Gnomovision comes with ABSOLUTELY NO WARRANTY; for details type `show w'.
This is free software, and you are welcome to redistribute it
under certain conditions; type `show c' for details.
The hypothetical commands `show w' and `show c' should show the appropriate
parts of the General Public License. Of course, the commands you use may
be called something other than `show w' and `show c'; they could even be
mouse-clicks or menu items--whatever suits your program.
You should also get your employer (if you work as a programmer) or your
school, if any, to sign a "copyright disclaimer" for the program, if
necessary. Here is a sample; alter the names:
Yoyodyne, Inc., hereby disclaims all copyright interest in the program
`Gnomovision' (which makes passes at compilers) written by James Hacker.
<signature of Ty Coon>, 1 April 1989
Ty Coon, President of Vice
This General Public License does not permit incorporating your program into
proprietary programs. If your program is a subroutine library, you may
consider it more useful to permit linking proprietary applications with the
library. If this is what you want to do, use the GNU Lesser General
Public License instead of this License.

105
ext-lib/bleualign/README.md Normal file
View File

@@ -0,0 +1,105 @@
Bleualign
=========
An MT-based sentence alignment tool
Copyright ⓒ 2010
Rico Sennrich <sennrich@cl.uzh.ch>
A project of the Computational Linguistics Group at the University of Zurich (http://www.cl.uzh.ch).
Project Homepage: http://github.com/rsennrich/bleualign
This program is free software; you can redistribute it and/or modify it under the terms of the GNU General Public License as published by the Free Software Foundation
GENERAL INFO
------------
Bleualign is a tool to align parallel texts (i.e. a text and its translation) on a sentence level.
Additionally to the source and target text, Bleualign requires an automatic translation of at least one of the texts.
The alignment is then performed on the basis of the similarity (modified BLEU score) between the source text sentences (translated into the target language) and the target text sentences.
See section PUBLICATIONS for more details.
Obtaining an automatic translation is up to the user. The only requirement is that the translation must correspond line-by-line to the source text (no line breaks inserted or removed).
REQUIREMENTS
------------
The software was developed on Linux using Python 2.6, but should also support newer versions of Python (including 3.X) and other platforms.
Please report any issues you encounter to sennrich@cl.uzh.ch
USAGE INSTRUCTIONS
------------------
The input and output formats of bleualign are one sentence per line.
A line which only contains .EOA is considered a hard delimiter (end of article).
Sentence alignment does not cross these delimiters: reliable delimiters improve speed and performance, wrong ones will seriously degrade performance.
Given the files sourcetext.txt, targettext.txt and sourcetranslation.txt (the latter being sentence-aligned with sourcetext.txt), a sample call is
./bleualign.py -s sourcetext.txt -t targettext.txt --srctotarget sourcetranslation.txt -o outputfile
It is also possible to provide several translations and/or translations in the other translation direction.
bleualign will run once per translation provided, the final output being the intersection of the individual runs (i.e. sentence pairs produced in each individual run).
./bleualign.py -s sourcetext.txt -t targettext.txt --srctotarget sourcetranslation1.txt --srctotarget sourcetranslation2.txt --targettosrc targettranslation1.txt -o outputfile
./bleualign.py -h will show more usage options
To facilitate batch processing multiple files, `batch_align.py` can be used.
python batch_align directory source_suffix target_suffix translation_suffix
example: given the directory `raw_files` with the files `0.de`, `0.fr` and `0.trans` and so on, (`0.trans` being the translation of `0.de` into the target language), then this command will align all files:
python batch_align.py raw_files de fr trans
This will produce the files `0.de.aligned` and `0.fr.aligned`
Input files are expected to use UTF-8 encoding.
USAGE AS PYTHON MODULE
----------------------
Bleualign works as stand-alone script, but can also be imported as a module other Python projects.
For code examples, see the example/ directory. If you want to know all options, you can see Aligner.default_options variable in bleualign/aligner.py.
To use Bleualign as a Python module, the package needs to be installed (from a local copy) with:
python setup.py install
The Bleualign package can also be installed directly from Github with:
pip install git+https://github.com/rsennrich/Bleualign.git
EVALUATION
---------
Two hand-aligned documents are provided with the repository for development and testing.
Evaluation is performed if you add the argument `-d` for the development set, and `-e` for the test set.
An example command for aligning the development set (one long document with 468/554 sentences in DE/FR):
./bleualign.py --source eval/eval1957.de --target eval/eval1957.fr --srctotarget eval/eval1957.europarlfull.fr -d
An example command for aligning the test set (7 documents, totalling 993/1011 sentences in DE/FR):
./bleualign.py --source eval/eval1989.de --target eval/eval1989.fr --srctotarget eval/eval1989.europarlfull.fr -e
PUBLICATIONS
------------
The algorithm is described in
Rico Sennrich, Martin Volk (2010):
MT-based Sentence Alignment for OCR-generated Parallel Texts. In: Proceedings of AMTA 2010, Denver, Colorado.
Rico Sennrich; Martin Volk (2011):
Iterative, MT-based sentence alignment of parallel texts. In: NODALIDA 2011, Nordic Conference of Computational Linguistics, Riga.
CONTACT
-------
For questions and feeback, please contact sennrich@cl.uzh.ch or use the GitHub repository.

15
ext-lib/bleualign/_bleualign.py Normal file
View File

@@ -0,0 +1,15 @@
#!/usr/bin/python
# -*- coding: utf-8 -*-
# Copyright © 2010 University of Zürich
# Author: Rico Sennrich <sennrich@cl.uzh.ch>
# For licensing information, see LICENSE
import sys
from command_utils import load_arguments
from bleualign.align import Aligner
if __name__ == '__main__':
options = load_arguments(sys.argv)
a = Aligner(options)
a.mainloop()

51
ext-lib/bleualign/batch_align.py Normal file
View File

@@ -0,0 +1,51 @@
#!/usr/bin/python
# -*- coding: utf-8 -*-
# Copyright: University of Zurich
# Author: Rico Sennrich
# script to allow batch-alignment of multiple files. No multiprocessing.
# syntax: python batch_align directory source_suffix target_suffix translation_suffix
#
# example: given the directory batch-test with the files 0.de, 0.fr and 0.trans, 1.de, 1.fr and 1.trans and so on,
# (0.trans being the translation of 0.de into the target language),
# then this command will align all files: python batch_align.py batch-test/ de fr trans
#
# output files will have ending source_suffix.aligned and target_suffix.aligned
import sys
import os
from bleualign.align import Aligner
if len(sys.argv) < 2:
sys.stderr.write('Usage: python batch_align.py job_file\n')
exit()
job_fn = sys.argv[1]
options = {}
options['factored'] = False
options['filter'] = None
options['filterthreshold'] = 90
options['filterlang'] = None
options['targettosrc'] = []
options['eval'] = None
options['galechurch'] = None
options['verbosity'] = 1
options['printempty'] = False
jobs = []
with open(job_fn, 'r', encoding="utf-8") as f:
for line in f:
if not line.startswith("#"):
jobs.append(line.strip())
for rec in jobs:
translation_document, source_document, target_document, out_document = rec.split("\t")
options['srcfile'] = source_document
options['targetfile'] = target_document
options['srctotarget'] = [translation_document]
options['output'] = out_document
a = Aligner(options)
a.mainloop()

110
ext-lib/bleualign/bleualign.py Normal file
View File

@@ -0,0 +1,110 @@
# 2021/11/27
# bfsujason@163.com
"""
Usage:
python ext-lib/bleualign/bleualign.py \
-m data/mac/test/meta_data.tsv \
-s data/mac/test/zh \
-t data/mac/test/en \
-o data/mac/test/auto
"""
import os
import sys
import time
import shutil
import argparse
def main():
parser = argparse.ArgumentParser(description='Sentence alignment using Bleualign')
parser.add_argument('-s', '--src', type=str, required=True, help='Source directory.')
parser.add_argument('-t', '--tgt', type=str, required=True, help='Target directory.')
parser.add_argument('-o', '--out', type=str, required=True, help='Output directory.')
parser.add_argument('-m', '--meta', type=str, required=True, help='Metadata file.')
parser.add_argument('--tok', action='store_true', help='Use tokenized source trans and target text.')
args = parser.parse_args()
make_dir(args.out)
jobs = create_jobs(args.meta, args.src, args.tgt, args.out, args.tok)
job_path = os.path.abspath(os.path.join(args.out, 'bleualign.job'))
write_jobs(jobs, job_path)
bleualign_bin = os.path.abspath('ext-lib/bleualign/batch_align.py')
run_bleualign(bleualign_bin, job_path)
convert_format(args.out)
def convert_format(dir):
for file in os.listdir(dir):
if file.endswith('-s'):
file_id = file.split('.')[0]
src = os.path.join(dir, file)
tgt = os.path.join(dir, file_id + '.align-t')
out = os.path.join(dir, file_id + '.align')
_convert_format(src, tgt, out)
os.unlink(src)
os.unlink(tgt)
def _convert_format(src, tgt, path):
src_align = read_alignment(src)
tgt_align = read_alignment(tgt)
with open(path, 'wt', encoding='utf-8') as f:
for x, y in zip(src_align, tgt_align):
f.write("{}:{}\n".format(x,y))
def read_alignment(file):
alignment = []
with open(file, 'rt', encoding='utf-8') as f:
for line in f:
line = line.strip()
alignment.append([int(x) for x in line.split(',')])
return alignment
def run_bleualign(bin, job):
cmd = "python {} {}".format(bin, job)
os.system(cmd)
os.unlink(job)
def write_jobs(jobs, path):
jobs = '\n'.join(jobs)
with open(path, 'wt', encoding='utf-8') as f:
f.write(jobs)
def create_jobs(meta, src, tgt, out, is_tok):
jobs = []
fns = get_fns(meta)
for file in fns:
src_path = os.path.abspath(os.path.join(src, file))
trans_path = os.path.abspath(os.path.join(src, file + '.trans'))
if is_tok:
tgt_path = os.path.abspath(os.path.join(tgt, file + '.tok'))
else:
tgt_path = os.path.abspath(os.path.join(tgt, file))
out_path = os.path.abspath(os.path.join(out, file + '.align'))
jobs.append('\t'.join([trans_path, src_path, tgt_path, out_path]))
return jobs
def get_fns(meta):
fns = []
with open(meta, 'rt', encoding='utf-8') as f:
next(f) # skip header
for line in f:
recs = line.strip().split('\t')
fns.append(recs[0])
return fns
def make_dir(path):
if os.path.isdir(path):
shutil.rmtree(path)
os.makedirs(path, exist_ok=True)
if __name__ == '__main__':
t_0 = time.time()
main()
print("It takes {:.3f} seconds to align all the sentences.".format(time.time() - t_0))

0
ext-lib/bleualign/bleualign/__init__.py Normal file
View File

1183
ext-lib/bleualign/bleualign/align.py Normal file
View File

File diff suppressed because it is too large Load Diff

205
ext-lib/bleualign/bleualign/gale_church.py Normal file
View File

@@ -0,0 +1,205 @@
# -*- coding: utf-8 -*-
import math
# Based on Gale & Church 1993,
# "A Program for Aligning Sentences in Bilingual Corpora"
infinity = float("inf")
def erfcc(x):
"""Complementary error function."""
z = abs(x)
t = 1 / (1 + 0.5 * z)
r = t * math.exp(-z * z -
1.26551223 + t *
(1.00002368 + t *
(.37409196 + t *
(.09678418 + t *
(-.18628806 + t *
(.27886807 + t *
(-1.13520398 + t *
(1.48851587 + t *
(-.82215223 + t * .17087277)))))))))
if (x >= 0.):
return r
else:
return 2. - r
def norm_cdf(x):
"""Return the area under the normal distribution from M{-∞..x}."""
return 1 - 0.5 * erfcc(x / math.sqrt(2))
class LanguageIndependent(object):
# These are the language-independent probabilities and parameters
# given in Gale & Church
# for the computation, l_1 is always the language with less characters
PRIORS = {
(1, 0): 0.0099,
(0, 1): 0.0099,
(1, 1): 0.89,
(2, 1): 0.089,
(1, 2): 0.089,
(2, 2): 0.011,
}
AVERAGE_CHARACTERS = 1
VARIANCE_CHARACTERS = 6.8
def trace(backlinks, source, target):
links = set()
pos = (len(source) - 1, len(target) - 1)
#while pos != (-1, -1):
while pos[0] != -1 and pos[1] != -1:
#print(pos)
#print(backlinks)
#print(backlinks[pos])
s, t = backlinks[pos]
for i in range(s):
for j in range(t):
links.add((pos[0] - i, pos[1] - j))
pos = (pos[0] - s, pos[1] - t)
return links
def align_probability(i, j, source_sentences, target_sentences, alignment, params):
"""Returns the probability of the two sentences C{source_sentences[i]}, C{target_sentences[j]}
being aligned with a specific C{alignment}.
@param i: The offset of the source sentence.
@param j: The offset of the target sentence.
@param source_sentences: The list of source sentence lengths.
@param target_sentences: The list of target sentence lengths.
@param alignment: The alignment type, a tuple of two integers.
@param params: The sentence alignment parameters.
@returns: The probability of a specific alignment between the two sentences, given the parameters.
"""
l_s = sum(source_sentences[i - offset] for offset in range(alignment[0]))
l_t = sum(target_sentences[j - offset] for offset in range(alignment[1]))
try:
# actually, the paper says l_s * params.VARIANCE_CHARACTERS, this is based on the C
# reference implementation. With l_s in the denominator, insertions are impossible.
m = (l_s + l_t / params.AVERAGE_CHARACTERS) / 2
delta = (l_t - l_s * params.AVERAGE_CHARACTERS) / math.sqrt(m * params.VARIANCE_CHARACTERS)
except ZeroDivisionError:
delta = infinity
return 2 * (1 - norm_cdf(abs(delta))) * params.PRIORS[alignment]
def align_blocks(source_sentences, target_sentences, params = LanguageIndependent):
"""Creates the sentence alignment of two blocks of texts (usually paragraphs).
@param source_sentences: The list of source sentence lengths.
@param target_sentences: The list of target sentence lengths.
@param params: the sentence alignment parameters.
@return: The sentence alignments, a list of index pairs.
"""
alignment_types = list(params.PRIORS.keys())
# there are always three rows in the history (with the last of them being filled)
# and the rows are always |target_text| + 2, so that we never have to do
# boundary checks
D = [(len(target_sentences) + 2) * [0] for x in range(2)]
# for the first sentence, only substitution, insertion or deletion are
# allowed, and they are all equally likely ( == 1)
D.append([0, 1])
try:
D[-2][1] = 1
D[-2][2] = 1
except:
pass
backlinks = {}
for i in range(len(source_sentences)):
for j in range(len(target_sentences)):
m = []
for a in alignment_types:
k = D[-(1 + a[0])][j + 2 - a[1]]
if k > 0:
p = k * \
align_probability(i, j, source_sentences, target_sentences, a, params)
m.append((p, a))
if len(m) > 0:
v = max(m)
backlinks[(i, j)] = v[1]
D[-1].append(v[0])
else:
backlinks[(i, j)] = (1, 1)
D[-1].append(0)
D.pop(0)
D.append([0, 0])
return trace(backlinks, source_sentences, target_sentences)
def align_texts(source_blocks, target_blocks, params = LanguageIndependent):
"""Creates the sentence alignment of two texts.
Texts can consist of several blocks. Block boundaries cannot be crossed by sentence
alignment links.
Each block consists of a list that contains the lengths (in characters) of the sentences
in this block.
@param source_blocks: The list of blocks in the source text.
@param target_blocks: The list of blocks in the target text.
@param params: the sentence alignment parameters.
@returns: A list of sentence alignment lists
"""
if len(source_blocks) != len(target_blocks):
raise ValueError("Source and target texts do not have the same number of blocks.")
return [align_blocks(source_block, target_block, params)
for source_block, target_block in zip(source_blocks, target_blocks)]
def split_at(it, split_value):
"""Splits an iterator C{it} at values of C{split_value}.
Each instance of C{split_value} is swallowed. The iterator produces
subiterators which need to be consumed fully before the next subiterator
can be used.
"""
def _chunk_iterator(first):
v = first
while v != split_value:
yield v
v = next(it)
while True:
yield _chunk_iterator(next(it))
def parse_token_stream(stream, soft_delimiter, hard_delimiter):
"""Parses a stream of tokens and splits it into sentences (using C{soft_delimiter} tokens)
and blocks (using C{hard_delimiter} tokens) for use with the L{align_texts} function.
"""
return [
[sum(len(token) for token in sentence_it)
for sentence_it in split_at(block_it, soft_delimiter)]
for block_it in split_at(stream, hard_delimiter)]
if __name__ == "__main__":
import sys
from contextlib import nested
with nested(open(sys.argv[1], "r"), open(sys.argv[2], "r")) as (s, t):
source = parse_token_stream((l.strip() for l in s), ".EOS", ".EOP")
target = parse_token_stream((l.strip() for l in t), ".EOS", ".EOP")
print((align_texts(source, target)))

146
ext-lib/bleualign/bleualign/score.py Normal file
View File

@@ -0,0 +1,146 @@
#!/usr/bin/python
# -*- coding: utf-8 -*-
#File originally part of moses package: http://www.statmt.org/moses/ (as bleu.py)
#Stripped of unused code to reduce number of libraries used
# $Id$
'''Provides:
cook_refs(refs, n=4): Transform a list of reference sentences as strings into a form usable by cook_test().
cook_test(test, refs, n=4): Transform a test sentence as a string (together with the cooked reference sentences) into a form usable by score_cooked().
score_cooked(alltest, n=4): Score a list of cooked test sentences.
score_set(s, testid, refids, n=4): Interface with dataset.py; calculate BLEU score of testid against refids.
The reason for breaking the BLEU computation into three phases cook_refs(), cook_test(), and score_cooked() is to allow the caller to calculate BLEU scores for multiple test sets as efficiently as possible.
'''
from __future__ import division, print_function
import sys, math, re, xml.sax.saxutils
# Added to bypass NIST-style pre-processing of hyp and ref files -- wade
nonorm = 0
preserve_case = False
eff_ref_len = "shortest"
normalize1 = [
('<skipped>', ''), # strip "skipped" tags
(r'-\n', ''), # strip end-of-line hyphenation and join lines
(r'\n', ' '), # join lines
# (r'(\d)\s+(?=\d)', r'\1'), # join digits
]
normalize1 = [(re.compile(pattern), replace) for (pattern, replace) in normalize1]
normalize2 = [
(r'([\{-\~\[-\` -\&\(-\+\:-\@\/])',r' \1 '), # tokenize punctuation. apostrophe is missing
(r'([^0-9])([\.,])',r'\1 \2 '), # tokenize period and comma unless preceded by a digit
(r'([\.,])([^0-9])',r' \1 \2'), # tokenize period and comma unless followed by a digit
(r'([0-9])(-)',r'\1 \2 ') # tokenize dash when preceded by a digit
]
normalize2 = [(re.compile(pattern), replace) for (pattern, replace) in normalize2]
#combine normalize2 into a single regex.
normalize3 = re.compile(r'([\{-\~\[-\` -\&\(-\+\:-\@\/])|(?:(?<![0-9])([\.,]))|(?:([\.,])(?![0-9]))|(?:(?<=[0-9])(-))')
def normalize(s):
'''Normalize and tokenize text. This is lifted from NIST mteval-v11a.pl.'''
# Added to bypass NIST-style pre-processing of hyp and ref files -- wade
if (nonorm):
return s.split()
try:
s.split()
except:
s = " ".join(s)
# language-independent part:
for (pattern, replace) in normalize1:
s = re.sub(pattern, replace, s)
s = xml.sax.saxutils.unescape(s, {'&quot;':'"'})
# language-dependent part (assuming Western languages):
s = " %s " % s
if not preserve_case:
s = s.lower() # this might not be identical to the original
return [tok for tok in normalize3.split(s) if tok and tok != ' ']
def count_ngrams(words, n=4):
counts = {}
for k in range(1,n+1):
for i in range(len(words)-k+1):
ngram = tuple(words[i:i+k])
counts[ngram] = counts.get(ngram, 0)+1
return counts
def cook_refs(refs, n=4):
'''Takes a list of reference sentences for a single segment
and returns an object that encapsulates everything that BLEU
needs to know about them.'''
refs = [normalize(ref) for ref in refs]
maxcounts = {}
for ref in refs:
counts = count_ngrams(ref, n)
for (ngram,count) in list(counts.items()):
maxcounts[ngram] = max(maxcounts.get(ngram,0), count)
return ([len(ref) for ref in refs], maxcounts)
def cook_ref_set(ref, n=4):
'''Takes a reference sentences for a single segment
and returns an object that encapsulates everything that BLEU
needs to know about them. Also provides a set cause bleualign wants it'''
ref = normalize(ref)
counts = count_ngrams(ref, n)
return (len(ref), counts, frozenset(counts))
def cook_test(test, args, n=4):
'''Takes a test sentence and returns an object that
encapsulates everything that BLEU needs to know about it.'''
reflens, refmaxcounts = args
test = normalize(test)
result = {}
result["testlen"] = len(test)
# Calculate effective reference sentence length.
if eff_ref_len == "shortest":
result["reflen"] = min(reflens)
elif eff_ref_len == "average":
result["reflen"] = float(sum(reflens))/len(reflens)
elif eff_ref_len == "closest":
min_diff = None
for reflen in reflens:
if min_diff is None or abs(reflen-len(test)) < min_diff:
min_diff = abs(reflen-len(test))
result['reflen'] = reflen
result["guess"] = [max(len(test)-k+1,0) for k in range(1,n+1)]
result['correct'] = [0]*n
counts = count_ngrams(test, n)
for (ngram, count) in list(counts.items()):
result["correct"][len(ngram)-1] += min(refmaxcounts.get(ngram,0), count)
return result
def score_cooked(allcomps, n=4):
totalcomps = {'testlen':0, 'reflen':0, 'guess':[0]*n, 'correct':[0]*n}
for comps in allcomps:
for key in ['testlen','reflen']:
totalcomps[key] += comps[key]
for key in ['guess','correct']:
for k in range(n):
totalcomps[key][k] += comps[key][k]
logbleu = 0.0
for k in range(n):
if totalcomps['correct'][k] == 0:
return 0.0
#log.write("%d-grams: %f\n" % (k,float(totalcomps['correct'][k])/totalcomps['guess'][k]))
logbleu += math.log(totalcomps['correct'][k])-math.log(totalcomps['guess'][k])
logbleu /= float(n)
#log.write("Effective reference length: %d test length: %d\n" % (totalcomps['reflen'], totalcomps['testlen']))
logbleu += min(0,1-float(totalcomps['reflen'])/totalcomps['testlen'])
return math.exp(logbleu)

191
ext-lib/bleualign/bleualign/utils.py Normal file
View File

@@ -0,0 +1,191 @@
#!/usr/bin/python
# -*- coding: utf-8 -*-
# Copyright: University of Zurich
# Author: Rico Sennrich
# For licensing information, see LICENSE
# Evaluation functions for Bleualign
from __future__ import division
from operator import itemgetter
def evaluate(options, testalign, goldalign, log_function):
goldalign = [(tuple(src),tuple(target)) for src,target in goldalign]
results = {}
paircounts = {}
for pair in [(len(srclist),len(targetlist)) for srclist,targetlist in goldalign]:
paircounts[pair] = paircounts.get(pair,0) + 1
pairs_normalized = {}
for pair in paircounts:
pairs_normalized[pair] = (paircounts[pair],paircounts[pair] / float(len(goldalign)))
log_function('\ngold alignment frequencies\n')
for aligntype,(abscount,relcount) in sorted(list(pairs_normalized.items()),key=itemgetter(1),reverse=True):
log_function(aligntype,end='')
log_function(' - ',end='')
log_function(abscount,end='')
log_function(' ('+str(relcount)+')')
log_function('\ntotal recall: ',end='')
log_function(str(len(goldalign)) + ' pairs in gold')
(tpstrict,fnstrict,tplax,fnlax) = recall((0,0),goldalign,[i[0] for i in testalign],log_function)
results['recall'] = (tpstrict,fnstrict,tplax,fnlax)
for aligntype in set([i[1] for i in testalign]):
testalign_bytype = []
for i in testalign:
if i[1] == aligntype:
testalign_bytype.append(i)
log_function('precision for alignment type ' + str(aligntype) + ' ( ' + str(len(testalign_bytype)) + ' alignment pairs)')
precision(goldalign,testalign_bytype,log_function)
log_function('\ntotal precision:',end='')
log_function(str(len(testalign)) + ' alignment pairs found')
(tpstrict,fpstrict,tplax,fplax) = precision(goldalign,testalign,log_function)
results['precision'] = (tpstrict,fpstrict,tplax,fplax)
return results
def precision(goldalign, testalign, log_function):
tpstrict=0
tplax=0
fpstrict=0
fplax=0
for (src,target) in [i[0] for i in testalign]:
if (src,target) == ((),()):
continue
if (src,target) in goldalign:
tpstrict +=1
tplax += 1
else:
srcset, targetset = set(src), set(target)
for srclist,targetlist in goldalign:
#lax condition: hypothesis and gold alignment only need to overlap
if srcset.intersection(set(srclist)) and targetset.intersection(set(targetlist)):
fpstrict +=1
tplax += 1
break
else:
fpstrict +=1
fplax +=1
log_function('false positive: ',2)
log_function((src,target),2)
if tpstrict+fpstrict > 0:
log_function('precision strict: ',end='')
log_function((tpstrict/float(tpstrict+fpstrict)))
log_function('precision lax: ',end='')
log_function((tplax/float(tplax+fplax)))
log_function('')
else:
log_function('nothing to find')
return tpstrict,fpstrict,tplax,fplax
def recall(aligntype, goldalign, testalign, log_function):
srclen,targetlen = aligntype
if srclen == 0 and targetlen == 0:
gapdists = [(0,0) for i in goldalign]
elif srclen == 0 or targetlen == 0:
log_function('nothing to find')
return
else:
gapdists = [(len(srclist),len(targetlist)) for srclist,targetlist in goldalign]
tpstrict=0
tplax=0
fnstrict=0
fnlax=0
for i,pair in enumerate(gapdists):
if aligntype == pair:
(srclist,targetlist) = goldalign[i]
if not srclist or not targetlist:
continue
elif (srclist,targetlist) in testalign:
tpstrict +=1
tplax +=1
else:
srcset, targetset = set(srclist), set(targetlist)
for src,target in testalign:
#lax condition: hypothesis and gold alignment only need to overlap
if srcset.intersection(set(src)) and targetset.intersection(set(target)):
tplax +=1
fnstrict+=1
break
else:
fnstrict+=1
fnlax+=1
log_function('not found: ',2),
log_function(goldalign[i],2)
if tpstrict+fnstrict>0:
log_function('recall strict: '),
log_function((tpstrict/float(tpstrict+fnstrict)))
log_function('recall lax: '),
log_function((tplax/float(tplax+fnlax)))
log_function('')
else:
log_function('nothing to find')
return tpstrict,fnstrict,tplax,fnlax
def finalevaluation(results, log_function):
recall_value = [0,0,0,0]
precision_value = [0,0,0,0]
for i,k in list(results.items()):
for m,j in enumerate(recall_value):
recall_value[m] = j+ k['recall'][m]
for m,j in enumerate(precision_value):
precision_value[m] = j+ k['precision'][m]
try:
pstrict = (precision_value[0]/float(precision_value[0]+precision_value[1]))
except ZeroDivisionError:
pstrict = 0
try:
plax =(precision_value[2]/float(precision_value[2]+precision_value[3]))
except ZeroDivisionError:
plax = 0
try:
rstrict= (recall_value[0]/float(recall_value[0]+recall_value[1]))
except ZeroDivisionError:
rstrict = 0
try:
rlax=(recall_value[2]/float(recall_value[2]+recall_value[3]))
except ZeroDivisionError:
rlax = 0
if (pstrict+rstrict) == 0:
fstrict = 0
else:
fstrict=2*(pstrict*rstrict)/(pstrict+rstrict)
if (plax+rlax) == 0:
flax=0
else:
flax=2*(plax*rlax)/(plax+rlax)
log_function('\n=========================\n')
log_function('total results:')
log_function('recall strict: ',end='')
log_function(rstrict)
log_function('recall lax: ',end='')
log_function(rlax)
log_function('')
log_function('precision strict: ',end='')
log_function(pstrict)
log_function('precision lax: '),
log_function(plax)
log_function('')
log_function('f1 strict: ',end='')
log_function(fstrict)
log_function('f1 lax: ',end='')
log_function(flax)
log_function('')

158
ext-lib/bleualign/command_utils.py Normal file
View File

@@ -0,0 +1,158 @@
#!/usr/bin/python
# -*- coding: utf-8 -*-
# Copyright © 2010 University of Zürich
# Author: Rico Sennrich <sennrich@cl.uzh.ch>
# For licensing information, see LICENSE
from __future__ import division, print_function
import sys
import os
import getopt
def usage():
bold = "\033[1m"
reset = "\033[0;0m"
italic = "\033[3m"
print('\n\t All files need to be one sentence per line and have .EOA as a hard delimiter. --source, --target and --output are mandatory arguments, the others are optional.')
print('\n\t' + bold +'--help' + reset + ', ' + bold +'-h' + reset)
print('\t\tprint usage information\n')
print('\t' + bold +'--source' + reset + ', ' + bold +'-s' + reset + ' file')
print('\t\tSource language text.')
print('\t' + bold +'--target' + reset + ', ' + bold +'-t' + reset + ' file')
print('\t\tTarget language text.')
print('\t' + bold +'--output' + reset + ', ' + bold +'-o' + reset + ' filename')
print('\t\tOutput file: Will create ' + 'filename' + '-s and ' + 'filename' + '-t')
print('\n\t' + bold +'--srctotarget' + reset + ' file')
print('\t\tTranslation of source language text to target language. Needs to be sentence-aligned with source language text.')
print('\t' + bold +'--targettosrc' + reset + ' file')
print('\t\tTranslation of target language text to source language. Needs to be sentence-aligned with target language text.')
print('\n\t' + bold +'--factored' + reset)
print('\t\tSource and target text can be factored (as defined by moses: | as separator of factors, space as word separator). Only first factor will be used for BLEU score.')
print('\n\t' + bold +'--filter' + reset + ', ' + bold +'-f' + reset + ' option')
print('\t\tFilters output. Possible options:')
print('\t\t' + bold +'sentences' + reset + '\tevaluate each sentence and filter on a per-sentence basis')
print('\t\t' + bold +'articles' + reset + '\tevaluate each article and filter on a per-article basis')
print('\n\t' + bold +'--filterthreshold' + reset + ' int')
print('\t\tFilters output to best XX percent. (Default: 90). Only works if --filter is set.')
print('\t' + bold +'--bleuthreshold' + reset + ' float')
print('\t\tFilters out sentence pairs with sentence-level BLEU score < XX (in range from 0 to 1). (Default: 0). Only works if --filter is set.')
print('\t' + bold +'--filterlang' + reset)
print('\t\tFilters out sentences/articles for which BLEU score between source and target is higher than that between translation and target (usually means source and target are in same language). Only works if --filter is set.')
print('\n\t' + bold +'--bleu_n' + reset + ' int')
print('\t\tConsider n-grams up to size n for BLEU. Default 2.')
print('\t' + bold +'--bleu_charlevel' + reset)
print('\t\tPerform BLEU on charcter-level (recommended for continuous script language; also consider increasing bleu_n).')
print('\n\t' + bold +'--galechurch' + reset)
print('\t\tAlign the bitext using Gale and Church\'s algorithm (without BLEU comparison).')
print('\t' + bold +'--printempty' + reset)
print('\t\tAlso write unaligned sentences to file. By default, they are discarded.')
print('\t' + bold +'--verbosity' + reset + ', ' + bold +'-v' + reset + ' int')
print('\t\tVerbosity. Choose amount of debugging output. Default value 1; choose 0 for (mostly) quiet mode, 2 for verbose output')
print('\t' + bold +'--processes' + reset + ', ' + bold +'-p' + reset + ' int')
print('\t\tNumber of parallel processes. Documents are split across available processes. Default: 4.')
def load_arguments(sysargv):
try:
opts, args = getopt.getopt(sysargv[1:], "def:ho:s:t:v:p:", ["factored", "filter=", "filterthreshold=", "bleuthreshold=", "filterlang", "printempty", "deveval","eval", "help", "bleu_n=", "bleu_charlevel", "galechurch", "output=", "source=", "target=", "srctotarget=", "targettosrc=", "verbosity=", "printempty=", "processes="])
except getopt.GetoptError as err:
# print help information and exit:
print(str(err)) # will print something like "option -a not recognized"
usage()
sys.exit(2)
options = {}
options['srcfile'] = None
options['targetfile'] = None
options['output'] = None
options['srctotarget'] = []
options['targettosrc'] = []
options['processes'] = 4
bold = "\033[1m"
reset = "\033[0;0m"
project_path = os.path.dirname(os.path.abspath(__file__))
for o, a in opts:
if o in ("-h", "--help"):
usage()
sys.exit()
elif o in ("-e", "--eval"):
options['srcfile'] = os.path.join(project_path,'eval','eval1989.de')
options['targetfile'] = os.path.join(project_path,'eval','eval1989.fr')
from eval import goldeval
goldalign = [None] * len(goldeval.gold1990map)
for index, data in list(goldeval.gold1990map.items()):
goldalign[index] = goldeval.gold[data]
options['eval'] = goldalign
elif o in ("-d", "--deveval"):
options['srcfile'] = os.path.join(project_path,'eval','eval1957.de')
options['targetfile'] = os.path.join(project_path,'eval','eval1957.fr')
from eval import golddev
goldalign = [golddev.goldalign]
options['eval'] = goldalign
elif o in ("-o", "--output"):
options['output'] = a
elif o == "--factored":
options['factored'] = True
elif o in ("-f", "--filter"):
if a in ['sentences','articles']:
options['filter'] = a
else:
print('\nERROR: Valid values for option ' + bold + '--filter'+ reset +' are '+ bold +'sentences '+ reset +'and ' + bold +'articles'+ reset +'.')
usage()
sys.exit(2)
elif o == "--filterthreshold":
options['filterthreshold'] = float(a)
elif o == "--bleuthreshold":
options['bleuthreshold'] = float(a)
elif o == "--filterlang":
options['filterlang'] = True
elif o == "--galechurch":
options['galechurch'] = True
elif o == "--bleu_n":
options['bleu_ngrams'] = int(a)
elif o == "--bleu_charlevel":
options['bleu_charlevel'] = True
elif o in ("-s", "--source"):
if not 'eval' in options:
options['srcfile'] = a
elif o in ("-t", "--target"):
if not 'eval' in options:
options['targetfile'] = a
elif o == "--srctotarget":
if a == '-':
options['no_translation_override'] = True
else:
options['srctotarget'].append(a)
elif o == "--targettosrc":
options['targettosrc'].append(a)
elif o == "--printempty":
options['printempty'] = True
elif o in ("-v", "--verbosity"):
global loglevel
loglevel = int(a)
options['loglevel'] = int(a)
options['verbosity'] = int(a)
elif o in ("-p", "--processes"):
options['num_processes'] = int(a)
else:
assert False, "unhandled option"
if not options['output']:
print('WARNING: Output not specified. Just printing debugging output.',0)
if not options['srcfile']:
print('\nERROR: Source file not specified.')
usage()
sys.exit(2)
if not options['targetfile']:
print('\nERROR: Target file not specified.')
usage()
sys.exit(2)
if options['targettosrc'] and not options['srctotarget']:
print('\nWARNING: Only --targettosrc specified, but expecting at least one --srctotarget. Please swap source and target side.')
sys.exit(2)
if not options['srctotarget'] and not options['targettosrc']\
and 'no_translation_override' not in options:
print("ERROR: no translation available: BLEU scores can be computed between the source and target text, but this is not the intended usage of Bleualign and may result in poor performance! If you're *really* sure that this is what you want, use the option '--srctotarget -'")
sys.exit(2)
return options

42
ext-lib/bleualign/setup.py Normal file
View File

@@ -0,0 +1,42 @@
# -*- coding: utf-8 -*-
import os
import setuptools
def read_file(filename):
return open(os.path.join(os.path.dirname(__file__), filename)).read()
setuptools.setup(
name = 'bleualign',
version = '0.1.1',
description = 'An MT-based sentence alignment tool',
long_description = read_file('README.md'),
author = 'Rico Sennrich',
author_email = 'sennrich@cl.uzh.ch',
url = 'https://github.com/rsennrich/Bleualign',
download_url = 'https://github.com/rsennrich/Bleualign',
keywords = [
'Sentence Alignment',
'Natural Language Processing',
'Statistical Machine Translation',
'BLEU',
],
classifiers = [
# which Development Status?
# 'Development Status :: 3 - Alpha',
'Development Status :: 4 - Beta',
# 'Development Status :: 5 - Production/Stable',
'License :: OSI Approved :: GNU General Public License v2 (GPLv2)',
'Operating System :: OS Independent',
'Programming Language :: Python :: 2.6',
'Programming Language :: Python :: 2.7',
'Programming Language :: Python :: 3',
'Programming Language :: Python :: 3.2',
'Programming Language :: Python :: 3.3',
'Programming Language :: Python :: 3.4',
'Topic :: Scientific/Engineering',
'Topic :: Scientific/Engineering :: Information Analysis',
'Topic :: Text Processing',
'Topic :: Text Processing :: Linguistic',
],
packages = ['bleualign'],
)

224381
ext-lib/hunalign/ce.dic Normal file
View File

File diff suppressed because it is too large Load Diff

BIN
ext-lib/hunalign/cygwin1.dll Normal file
View File

Binary file not shown.

224381
ext-lib/hunalign/ec.dic Normal file
View File

File diff suppressed because it is too large Load Diff

0
ext-lib/hunalign/empty.dic Normal file
View File

BIN
ext-lib/hunalign/hunalign Normal file
View File

Binary file not shown.

BIN
ext-lib/hunalign/hunalign.exe Normal file
View File

Binary file not shown.

136
ext-lib/hunalign/hunalign.py Normal file
View File

@@ -0,0 +1,136 @@
# 2021/11/27
# bfsujason@163.com
"""
Usage:
python ext-lib/hunalign/hunalign.py \
-m data/mac/test/meta_data.tsv \
-s data/mac/test/zh \
-t data/mac/test/en \
-o data/mac/test/auto \
-d ec.dic
"""
import os
import time
import shutil
import platform
import argparse
def main():
parser = argparse.ArgumentParser(description='Sentence alignment using Hunalign')
parser.add_argument('-s', '--src', type=str, required=True, help='Source directory.')
parser.add_argument('-t', '--tgt', type=str, required=True, help='Target directory.')
parser.add_argument('-o', '--out', type=str, required=True, help='Output directory.')
parser.add_argument('-m', '--meta', type=str, required=True, help='Metadata file.')
parser.add_argument('-d', '--dic', type=str, help='Dictionary file.')
args = parser.parse_args()
make_dir(args.out)
jobs = create_jobs(args.meta, args.src, args.tgt, args.out)
job_path = os.path.abspath(os.path.join(args.out, 'hunalign.job'))
write_jobs(jobs, job_path)
if args.dic:
hunalign_dic = os.path.abspath(os.path.join('ext-lib/hunalign', args.dic))
else:
hunalign_dic = os.path.abspath('ext-lib/hunalign/null.dic')
# check system OS
OS = platform.system()
if OS == 'Windows':
hunalign_bin = os.path.abspath('ext-lib/hunalign/hunalign.exe')
elif OS == 'Linux':
hunalign_bin = os.path.abspath('ext-lib/hunalign/hunalign')
print(hunalign_bin)
print(hunalign_dic)
print(job_path)
run_hunalign(hunalign_bin, hunalign_dic, job_path)
convert_format(args.out)
def convert_format(dir):
for file in sorted(os.listdir(dir)):
fp_in = os.path.join(dir, file)
fp_out = os.path.join(dir, file + '.align')
alignment = _convert_format(fp_in, fp_out)
write_alignment(alignment, fp_out)
os.unlink(fp_in)
def _convert_format(fp_in, fp_out):
src_id = -1
tgt_id = -1
alignment = []
with open(fp_in, 'rt', encoding='utf-8') as f:
for line in f:
line = line.strip(' \r\n')
items = line.split('\t');
if not items[0] and not items[1]:
continue
src_seg_len, src_seg_id = _parse_seg(items[0], src_id)
tgt_seg_len, tgt_seg_id = _parse_seg(items[1], tgt_id)
src_id += src_seg_len
tgt_id += tgt_seg_len
alignment.append((src_seg_id, tgt_seg_id))
return alignment
def write_alignment(alignment, fp_out):
with open(fp_out, 'wt', encoding='utf-8') as f:
for id in alignment:
f.write("{}:{}\n".format(id[0], id[1]))
def _parse_seg(seg, id):
seg_len = 0
seg_id = []
if seg:
sents = seg.split(' ~~~ ')
seg_len = len(sents)
seg_id = [id + x for x in range(1, seg_len+1)]
return seg_len, seg_id
def run_hunalign(bin, dic, job):
cmd = "{} -text -batch {} {}".format(bin, dic, job)
os.system(cmd)
os.unlink(job)
def write_jobs(jobs, path):
jobs = '\n'.join(jobs)
with open(path, 'wt', encoding='utf-8', newline='\n') as f:
f.write(jobs)
def create_jobs(meta, src, tgt, out):
jobs = []
fns = get_fns(meta)
for file in fns:
# using tokenized file
src_path = os.path.abspath(os.path.join(src, file + '.tok'))
tgt_path = os.path.abspath(os.path.join(tgt, file + '.tok'))
out_path = os.path.abspath(os.path.join(out, file))
jobs.append('\t'.join([src_path, tgt_path, out_path]))
return jobs
def get_fns(meta):
fns = []
with open(meta, 'rt', encoding='utf-8') as f:
next(f) # skip header
for line in f:
recs = line.strip().split('\t')
fns.append(recs[0])
return fns
def make_dir(path):
if os.path.isdir(path):
shutil.rmtree(path)
os.makedirs(path, exist_ok=True)
if __name__ == '__main__':
t_0 = time.time()
main()
print("It takes {:.3f} seconds to align all the sentences.".format(time.time() - t_0))

0
ext-lib/hunalign/null.dic Normal file
View File

0
ext-lib/hunalign/translate.txt Normal file
View File

9
ext-lib/vecalign/.gitignore vendored Normal file
View File

@@ -0,0 +1,9 @@
build/
dp_core.c*
dp_core.html
__pycache__/
.idea
*~
.pytest_cache/
venv/

202
ext-lib/vecalign/LICENSE Normal file
View File

@@ -0,0 +1,202 @@
Apache License
Version 2.0, January 2004
http://www.apache.org/licenses/
TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
1. Definitions.
"License" shall mean the terms and conditions for use, reproduction,
and distribution as defined by Sections 1 through 9 of this document.
"Licensor" shall mean the copyright owner or entity authorized by
the copyright owner that is granting the License.
"Legal Entity" shall mean the union of the acting entity and all
other entities that control, are controlled by, or are under common
control with that entity. For the purposes of this definition,
"control" means (i) the power, direct or indirect, to cause the
direction or management of such entity, whether by contract or
otherwise, or (ii) ownership of fifty percent (50%) or more of the
outstanding shares, or (iii) beneficial ownership of such entity.
"You" (or "Your") shall mean an individual or Legal Entity
exercising permissions granted by this License.
"Source" form shall mean the preferred form for making modifications,
including but not limited to software source code, documentation
source, and configuration files.
"Object" form shall mean any form resulting from mechanical
transformation or translation of a Source form, including but
not limited to compiled object code, generated documentation,
and conversions to other media types.
"Work" shall mean the work of authorship, whether in Source or
Object form, made available under the License, as indicated by a
copyright notice that is included in or attached to the work
(an example is provided in the Appendix below).
"Derivative Works" shall mean any work, whether in Source or Object
form, that is based on (or derived from) the Work and for which the
editorial revisions, annotations, elaborations, or other modifications
represent, as a whole, an original work of authorship. For the purposes
of this License, Derivative Works shall not include works that remain
separable from, or merely link (or bind by name) to the interfaces of,
the Work and Derivative Works thereof.
"Contribution" shall mean any work of authorship, including
the original version of the Work and any modifications or additions
to that Work or Derivative Works thereof, that is intentionally
submitted to Licensor for inclusion in the Work by the copyright owner
or by an individual or Legal Entity authorized to submit on behalf of
the copyright owner. For the purposes of this definition, "submitted"
means any form of electronic, verbal, or written communication sent
to the Licensor or its representatives, including but not limited to
communication on electronic mailing lists, source code control systems,
and issue tracking systems that are managed by, or on behalf of, the
Licensor for the purpose of discussing and improving the Work, but
excluding communication that is conspicuously marked or otherwise
designated in writing by the copyright owner as "Not a Contribution."
"Contributor" shall mean Licensor and any individual or Legal Entity
on behalf of whom a Contribution has been received by Licensor and
subsequently incorporated within the Work.
2. Grant of Copyright License. Subject to the terms and conditions of
this License, each Contributor hereby grants to You a perpetual,
worldwide, non-exclusive, no-charge, royalty-free, irrevocable
copyright license to reproduce, prepare Derivative Works of,
publicly display, publicly perform, sublicense, and distribute the
Work and such Derivative Works in Source or Object form.
3. Grant of Patent License. Subject to the terms and conditions of
this License, each Contributor hereby grants to You a perpetual,
worldwide, non-exclusive, no-charge, royalty-free, irrevocable
(except as stated in this section) patent license to make, have made,
use, offer to sell, sell, import, and otherwise transfer the Work,
where such license applies only to those patent claims licensable
by such Contributor that are necessarily infringed by their
Contribution(s) alone or by combination of their Contribution(s)
with the Work to which such Contribution(s) was submitted. If You
institute patent litigation against any entity (including a
cross-claim or counterclaim in a lawsuit) alleging that the Work
or a Contribution incorporated within the Work constitutes direct
or contributory patent infringement, then any patent licenses
granted to You under this License for that Work shall terminate
as of the date such litigation is filed.
4. Redistribution. You may reproduce and distribute copies of the
Work or Derivative Works thereof in any medium, with or without
modifications, and in Source or Object form, provided that You
meet the following conditions:
(a) You must give any other recipients of the Work or
Derivative Works a copy of this License; and
(b) You must cause any modified files to carry prominent notices
stating that You changed the files; and
(c) You must retain, in the Source form of any Derivative Works
that You distribute, all copyright, patent, trademark, and
attribution notices from the Source form of the Work,
excluding those notices that do not pertain to any part of
the Derivative Works; and
(d) If the Work includes a "NOTICE" text file as part of its
distribution, then any Derivative Works that You distribute must
include a readable copy of the attribution notices contained
within such NOTICE file, excluding those notices that do not
pertain to any part of the Derivative Works, in at least one
of the following places: within a NOTICE text file distributed
as part of the Derivative Works; within the Source form or
documentation, if provided along with the Derivative Works; or,
within a display generated by the Derivative Works, if and
wherever such third-party notices normally appear. The contents
of the NOTICE file are for informational purposes only and
do not modify the License. You may add Your own attribution
notices within Derivative Works that You distribute, alongside
or as an addendum to the NOTICE text from the Work, provided
that such additional attribution notices cannot be construed
as modifying the License.
You may add Your own copyright statement to Your modifications and
may provide additional or different license terms and conditions
for use, reproduction, or distribution of Your modifications, or
for any such Derivative Works as a whole, provided Your use,
reproduction, and distribution of the Work otherwise complies with
the conditions stated in this License.
5. Submission of Contributions. Unless You explicitly state otherwise,
any Contribution intentionally submitted for inclusion in the Work
by You to the Licensor shall be under the terms and conditions of
this License, without any additional terms or conditions.
Notwithstanding the above, nothing herein shall supersede or modify
the terms of any separate license agreement you may have executed
with Licensor regarding such Contributions.
6. Trademarks. This License does not grant permission to use the trade
names, trademarks, service marks, or product names of the Licensor,
except as required for reasonable and customary use in describing the
origin of the Work and reproducing the content of the NOTICE file.
7. Disclaimer of Warranty. Unless required by applicable law or
agreed to in writing, Licensor provides the Work (and each
Contributor provides its Contributions) on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
implied, including, without limitation, any warranties or conditions
of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
PARTICULAR PURPOSE. You are solely responsible for determining the
appropriateness of using or redistributing the Work and assume any
risks associated with Your exercise of permissions under this License.
8. Limitation of Liability. In no event and under no legal theory,
whether in tort (including negligence), contract, or otherwise,
unless required by applicable law (such as deliberate and grossly
negligent acts) or agreed to in writing, shall any Contributor be
liable to You for damages, including any direct, indirect, special,
incidental, or consequential damages of any character arising as a
result of this License or out of the use or inability to use the
Work (including but not limited to damages for loss of goodwill,
work stoppage, computer failure or malfunction, or any and all
other commercial damages or losses), even if such Contributor
has been advised of the possibility of such damages.
9. Accepting Warranty or Additional Liability. While redistributing
the Work or Derivative Works thereof, You may choose to offer,
and charge a fee for, acceptance of support, warranty, indemnity,
or other liability obligations and/or rights consistent with this
License. However, in accepting such obligations, You may act only
on Your own behalf and on Your sole responsibility, not on behalf
of any other Contributor, and only if You agree to indemnify,
defend, and hold each Contributor harmless for any liability
incurred by, or claims asserted against, such Contributor by reason
of your accepting any such warranty or additional liability.
END OF TERMS AND CONDITIONS
APPENDIX: How to apply the Apache License to your work.
To apply the Apache License to your work, attach the following
boilerplate notice, with the fields enclosed by brackets "[]"
replaced with your own identifying information. (Don't include
the brackets!) The text should be enclosed in the appropriate
comment syntax for the file format. We also recommend that a
file or class name and description of purpose be included on the
same "printed page" as the copyright notice for easier
identification within third-party archives.
Copyright [yyyy] [name of copyright owner]
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License.

158
ext-lib/vecalign/README.md Normal file
View File

@@ -0,0 +1,158 @@
# Vecalign
Vecalign is an accurate sentence alignment algorithm which is fast even for very long documents.
In conjunction with [LASER](https://github.com/facebookresearch/LASER), Vecalign
works in about 100 languages (i.e. 100^2 language pairs),
without the need for a machine translation system or lexicon.
Vecalign uses similarity of multilingual sentence embeddings to judge the similarity of sentences.
![multilingual_sentence_embedding image](media/multilingual_sentence_embedding.png)
[image based on [this Facebook AI post](https://engineering.fb.com/ai-research/laser-multilingual-sentence-embeddings/)]
Vecalign uses an approximation to Dynamic Programming based on
[Fast Dynamic Time Warping](https://content.iospress.com/articles/intelligent-data-analysis/ida00303)
which is linear in time and space with respect to the number of sentences being aligned.
![dynamic_programing_approximation visualization](media/dynamic_programing_approximation.gif)
### License
Copyright 2019 Brian Thompson
Vecalign is released under the [Apache License, Version 2.0](LICENSE).
For convenience, the dev and test datasets from Bleualign are provided. Bleualign is Copyright 2010 Rico Sennrich and is released under the [GNU General Public License Version 2](bleualign_data/LICENSE)
### Build Vecalign
You will need python 3.6+ with numpy and cython. You can build an environment using conda as follows:
```
# Use latest conda
conda update conda -y
# Create conda environment
conda create --force -y --name vecalign python=3.7
# Activate new environment
source `conda info --base`/etc/profile.d/conda.sh # See: https://github.com/conda/conda/issues/7980
conda activate vecalign
# Install required packages
conda install -y -c anaconda cython
conda install -y -c anaconda numpy
```
Note that Vecalign contains cython code, but there is no need to build it manually as it is compiled automatically by [pyximport](https://github.com/cython/cython/tree/master/pyximport).
### Run Vecalign (using provided embeddings)
```
./vecalign.py --alignment_max_size 8 --src bleualign_data/dev.de --tgt bleualign_data/dev.fr \
--src_embed bleualign_data/overlaps.de bleualign_data/overlaps.de.emb \
--tgt_embed bleualign_data/overlaps.fr bleualign_data/overlaps.fr.emb
```
Alignments are written to stdout:
```
[0]:[0]:0.156006
[1]:[1]:0.160997
[2]:[2]:0.217155
[3]:[3]:0.361439
[4]:[4]:0.346332
[5]:[5]:0.211873
[6]:[6, 7, 8]:0.507506
[7]:[9]:0.252747
[8, 9]:[10, 11, 12]:0.139594
[10, 11]:[13]:0.273751
[12]:[14]:0.165397
[13]:[15, 16, 17]:0.436312
[14]:[18, 19, 20, 21]:0.734142
[]:[22]:0.000000
[]:[23]:0.000000
[]:[24]:0.000000
[]:[25]:0.000000
[15]:[26, 27, 28]:0.840094
...
```
The first two entries are the source and target sentence indexes for each alignment, respectively.
The third entry in each line is the sentence alignment cost computed by Vecalign.
Note that this cost includes normalization but does *not* include the penalties terms for containing more than one sentence.
Note that the alignment cost is set to zero for insertions/deletions.
Also note that the results may vary slightly due to randomness in the normalization.
To score against a gold alignment, use the "-g" flag.
Flags "-s", "-t", and "-g" can accept multiple arguments. This is primarily useful for scoring, as the output alignments will all be concatenated together in stdout. For example, to align and score the bleualign test set:
```
./vecalign.py --alignment_max_size 8 --src bleualign_data/test*.de --tgt bleualign_data/test*.fr \
--gold bleualign_data/test*.defr \
--src_embed bleualign_data/overlaps.de bleualign_data/overlaps.de.emb \
--tgt_embed bleualign_data/overlaps.fr bleualign_data/overlaps.fr.emb > /dev/null
```
Which should give you results that approximately match the Vecalign paper:
```
---------------------------------
| | Strict | Lax |
| Precision | 0.899 | 0.985 |
| Recall | 0.904 | 0.987 |
| F1 | 0.902 | 0.986 |
---------------------------------
```
Note: Run `./vecalign.py -h` for full sentence alignment usage and options.
For stand-alone scoring against a gold reference, see [score.py](score.py)
### Embed your own documents
The Vecalign repository contains overlap and embedding files for the Bluealign dev/test files.
This section shows how those files were made, as an example for running on new data.
Vecalign requires not only embeddings of sentences in each document,
but also embeddings of *concatenations* of consecutive sentences.
The embeddings of multiple, consecutive sentences are needed to consider 1-many, many-1, and many-many alignments.
To create a file containing all the sentence combinations in the dev and test files from Bleualign:
```
./overlap.py -i bleualign_data/dev.fr bleualign_data/test*.fr -o bleualign_data/overlaps.fr -n 10
./overlap.py -i bleualign_data/dev.de bleualign_data/test*.de -o bleualign_data/overlaps.de -n 10
```
Note: Run `./overlap.py -h` to see full set of embedding options.
`bleualign_data/overlaps.fr` and `bleualign_data/overlaps.de` are text files containing one or more sentences per line.
These files must then be embedded using a multilingual sentence embedder.
We recommend the [Language-Agnostic SEntence Representations (LASER)](https://github.com/facebookresearch/LASER)
toolkit from Facebook, as it has strong performance and comes with a pretrained model which works well in about 100 languages.
However, Vecalign should also work with other embedding methods as well. Embeddings should be provided as a binary file containing float32 values.
The following assumes LASER is installed and the LASER environmental variable has been set.
To embed the Bleualign files using LASER:
```
$LASER/tasks/embed/embed.sh bleualign_data/overlaps.fr fr bleualign_data/overlaps.fr.emb
$LASER/tasks/embed/embed.sh bleualign_data/overlaps.de de bleualign_data/overlaps.de.emb
```
Note that LASER will not overwrite an embedding file if it exsts, so you may need to run first `rm bleualign_data/overlaps.fr.emb bleualign_data/overlaps.de.emb`.
### Publications
If you use Vecalign, please cite our [paper](https://www.aclweb.org/anthology/D19-1136.pdf):
```
@inproceedings{thompson-koehn-2019-vecalign,
title = "{V}ecalign: Improved Sentence Alignment in Linear Time and Space",
author = "Thompson, Brian and Koehn, Philipp",
booktitle = "Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP)",
month = nov,
year = "2019",
address = "Hong Kong, China",
publisher = "Association for Computational Linguistics",
url = "https://www.aclweb.org/anthology/D19-1136",
doi = "10.18653/v1/D19-1136",
pages = "1342--1348",
}
```

0
ext-lib/vecalign/__init__.py Normal file
View File

148
ext-lib/vecalign/_vecalign.py Normal file
View File

@@ -0,0 +1,148 @@
#!/usr/bin/env python3
"""
Copyright 2019 Brian Thompson
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
https://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License.
"""
import argparse
import logging
import pickle
from math import ceil
from random import seed as seed
import numpy as np
logger = logging.getLogger('vecalign')
logger.setLevel(logging.WARNING)
logFormatter = logging.Formatter("%(asctime)s %(levelname)-5.5s %(message)s")
consoleHandler = logging.StreamHandler()
consoleHandler.setFormatter(logFormatter)
logger.addHandler(consoleHandler)
from dp_utils import make_alignment_types, print_alignments, read_alignments, \
read_in_embeddings, make_doc_embedding, vecalign
from score import score_multiple, log_final_scores
def _main():
# make runs consistent
seed(42)
np.random.seed(42)
parser = argparse.ArgumentParser('Sentence alignment using sentence embeddings and FastDTW',
formatter_class=argparse.ArgumentDefaultsHelpFormatter)
parser.add_argument('-s', '--src', type=str, nargs='+', required=True,
help='preprocessed source file to align')
parser.add_argument('-t', '--tgt', type=str, nargs='+', required=True,
help='preprocessed target file to align')
parser.add_argument('-g', '--gold_alignment', type=str, nargs='+', required=False,
help='preprocessed target file to align')
parser.add_argument('--src_embed', type=str, nargs=2, required=True,
help='Source embeddings. Requires two arguments: first is a text file, sencond is a binary embeddings file. ')
parser.add_argument('--tgt_embed', type=str, nargs=2, required=True,
help='Target embeddings. Requires two arguments: first is a text file, sencond is a binary embeddings file. ')
parser.add_argument('-a', '--alignment_max_size', type=int, default=4,
help='Searches for alignments up to size N-M, where N+M <= this value. Note that the the embeddings must support the requested number of overlaps')
parser.add_argument('-d', '--del_percentile_frac', type=float, default=0.2,
help='Deletion penalty is set to this percentile (as a fraction) of the cost matrix distribution. Should be between 0 and 1.')
parser.add_argument('-v', '--verbose', help='sets consle to logging.DEBUG instead of logging.WARN',
action='store_true')
parser.add_argument('--max_size_full_dp', type=int, default=300,
help='Maximum size N for which is is acceptable to run full N^2 dynamic programming.')
parser.add_argument('--costs_sample_size', type=int, default=20000,
help='Sample size to estimate costs distribution, used to set deletion penalty in conjunction with deletion_percentile.')
parser.add_argument('--num_samps_for_norm', type=int, default=100,
help='Number of samples used for normalizing embeddings')
parser.add_argument('--search_buffer_size', type=int, default=5,
help='Width (one side) of search buffer. Larger values makes search more likely to recover from errors but increases runtime.')
parser.add_argument('--debug_save_stack', type=str,
help='Write stack to pickle file for debug purposes')
args = parser.parse_args()
if len(args.src) != len(args.tgt):
raise Exception('number of source files must match number of target files')
if args.gold_alignment is not None:
if len(args.gold_alignment) != len(args.src):
raise Exception('number of gold alignment files, if provided, must match number of source and target files')
if args.verbose:
import logging
logger.setLevel(logging.INFO)
if args.alignment_max_size < 2:
logger.warning('Alignment_max_size < 2. Increasing to 2 so that 1-1 alignments will be considered')
args.alignment_max_size = 2
src_sent2line, src_line_embeddings = read_in_embeddings(args.src_embed[0], args.src_embed[1])
tgt_sent2line, tgt_line_embeddings = read_in_embeddings(args.tgt_embed[0], args.tgt_embed[1])
width_over2 = ceil(args.alignment_max_size / 2.0) + args.search_buffer_size
test_alignments = []
stack_list = []
for src_file, tgt_file in zip(args.src, args.tgt):
logger.info('Aligning src="%s" to tgt="%s"', src_file, tgt_file)
src_lines = open(src_file, 'rt', encoding="utf-8").readlines()
vecs0 = make_doc_embedding(src_sent2line, src_line_embeddings, src_lines, args.alignment_max_size)
tgt_lines = open(tgt_file, 'rt', encoding="utf-8").readlines()
vecs1 = make_doc_embedding(tgt_sent2line, tgt_line_embeddings, tgt_lines, args.alignment_max_size)
final_alignment_types = make_alignment_types(args.alignment_max_size)
logger.debug('Considering alignment types %s', final_alignment_types)
stack = vecalign(vecs0=vecs0,
vecs1=vecs1,
final_alignment_types=final_alignment_types,
del_percentile_frac=args.del_percentile_frac,
width_over2=width_over2,
max_size_full_dp=args.max_size_full_dp,
costs_sample_size=args.costs_sample_size,
num_samps_for_norm=args.num_samps_for_norm)
# write final alignments to stdout
print_alignments(stack[0]['final_alignments'], stack[0]['alignment_scores'])
test_alignments.append(stack[0]['final_alignments'])
stack_list.append(stack)
if args.gold_alignment is not None:
gold_list = [read_alignments(x) for x in args.gold_alignment]
res = score_multiple(gold_list=gold_list, test_list=test_alignments)
log_final_scores(res)
if args.debug_save_stack:
pickle.dump(stack_list, open(args.debug_save_stack, 'wb'))
if __name__ == '__main__':
_main()

411
ext-lib/vecalign/dp_core.pyx Normal file
View File

@@ -0,0 +1,411 @@
# cython: language_level=3
"""
Copyright 2019 Brian Thompson
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
https://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License.
"""
import numpy as np
cimport numpy as np
cimport cython
def make_x_y_offsets(alignment_types):
# alignment types for which we will precompute costs
# deletion/insertion is added later
for x, y in alignment_types:
assert (x > 0)
assert (y > 0)
x_offsets = np.array([x for x, y in alignment_types], dtype=np.int32) # MUST **NOT** INCLUDE (0,1), (1,0)
y_offsets = np.array([y for x, y in alignment_types], dtype=np.int32) # MUST **NOT** INCLUDE (0,1), (1,0)
return x_offsets, y_offsets
def make_dense_costs(np.ndarray[float, ndim=3] vecs0, # itput
np.ndarray[float, ndim=3] vecs1, # input
np.ndarray[float, ndim=2] norm0, # input
np.ndarray[float, ndim=2] norm1, # input
int offset0 = 0, # index into vecs0/norms0
int offset1 = 0, # index into vecs1/norms1
):
"""
Make a full N*M feature matrix. By default, makes 1-1 alignments,
can build others by specifying offset0, offset1 to index into
vecs0, norms0 and vecs1, norms1 respectivly.
"""
assert vecs0.shape[0] > offset0
assert vecs1.shape[0] > offset1
assert norm0.shape[0] > offset0
assert norm1.shape[0] > offset1
cdef int size0 = np.shape(vecs0)[1]
assert norm0.shape[1] == size0
cdef int size1 = np.shape(vecs1)[1]
assert norm1.shape[1] == size1
cdef int vecsize = np.shape(vecs0)[2]
assert vecs1.shape[2] == vecsize
cdef int xi, yi
cdef float sumx
cdef np.ndarray[float, ndim=2] costs = np.empty((size0, size1), dtype=np.float32)
for xi in range(size0):
for yi in range(size1):
sumx = 0.0
for jj in range(vecsize):
sumx += vecs0[offset0, xi, jj] * vecs1[offset1, yi, jj]
costs[xi, yi] = 2.0 * (1.0 - sumx) / (1e-6 + norm0[offset0, xi] + norm1[offset1, yi])
# normalize by alignment type
costs[xi, yi] = costs[xi, yi] * (offset0 + 1) * (offset1 + 1)
return costs
def dense_dp(np.ndarray[float, ndim=2] alignment_cost, float pen):
"""
Compute cost matrix (csum) and backpointers (bp)
from full 2-D 1-1 alignment costs matrix (alignment_cost)
"""
size0 = alignment_cost.shape[0]
size1 = alignment_cost.shape[1]
# csum and traceback matrix are both on nodes
# so they are +1 in each dimension compared to the jump costs matrix
# For anything being used in accumulation, use float64
cdef np.ndarray[double, ndim=2] csum = np.empty((size0 + 1, size1 + 1), dtype=np.float64)
cdef np.ndarray[int, ndim=2] bp = np.empty((size0 + 1, size1 + 1), dtype=np.int32)
# bp and csum are nodes,
# while alignment_cost is the cost of going between the nodes
# Size of nodes should be one larger than alignment costs
b0, b1 = np.shape(bp)
c0, c1 = np.shape(csum)
j0, j1 = np.shape(alignment_cost)
assert (b0 == c0 == j0 + 1)
assert (b1 == c1 == j1 + 1)
cdef int cmax = np.shape(csum)[1]
cdef int rmax = np.shape(csum)[0]
cdef int c, r
cdef double cost0, cost1, cost2
# initialize the all c-direction deletion path
for c in range(cmax):
csum[0, c] = c * pen
bp[0, c] = 1
# initialize the all r-direction deletion path
for r in range(rmax):
csum[r, 0] = r * pen
bp[r, 0] = 2
# Initial cost is 0.0
csum[0, 0] = 0.0 # noop
bp[0, 0] = 4 # should not matter
# Calculate the rest recursively
for c in range(1, cmax):
for r in range(1, rmax):
# alignment_cost indexes are off by 1 wrt
# csum/bp, since csum/bp are nodes
cost0 = csum[r - 1, c - 1] + alignment_cost[r - 1, c - 1]
cost1 = csum[r, c - 1] + pen
cost2 = csum[r - 1, c] + pen
csum[r, c] = cost0
bp[r, c] = 0
if cost1 < csum[r, c]:
csum[r, c] = cost1
bp[r, c] = 1
if cost2 < csum[r, c]:
csum[r, c] = cost2
bp[r, c] = 2
return csum, bp
def score_path(np.ndarray[int, ndim=1] xx,
np.ndarray[int, ndim=1] yy,
np.ndarray[float, ndim=1] norm1,
np.ndarray[float, ndim=1] norm2,
np.ndarray[float, ndim=2] vecs1,
np.ndarray[float, ndim=2] vecs2,
np.ndarray[float, ndim=1] out):
cdef int xi, yi, ii, jj
cdef float outx
cdef int lenxy = xx.shape[0]
cdef int vecsize = vecs1.shape[1]
for ii in range(lenxy):
xi = xx[ii]
yi = yy[ii]
outx = 0.0
for jj in range(vecsize):
outx += vecs1[xi, jj] * vecs2[yi, jj]
out[ii] = 2.0 * (1.0 - outx) / (norm1[xi] + norm2[yi])
# Bounds checking and wraparound slow things down by about 2x
# Division by 0 checking has minimal speed impact
@cython.boundscheck(False) # turn off bounds-checking for entire function
@cython.wraparound(False) # turn off negative index wrapping for entire function
@cython.cdivision(True) # use c-style division (no division-by-zero check)
def make_sparse_costs(np.ndarray[float, ndim=3] vecs0, # intput: num aligns X num sents X dim
np.ndarray[float, ndim=3] vecs1, # input
np.ndarray[float, ndim=2] norms0, # intput: num aligns X num sents
np.ndarray[float, ndim=2] norms1, # input
x_y_path,
alignment_types,
int width_over2):
"""
Make features for DP, *for lines running across approximate path*, *for each alignment type*
x_offsets, y_offsets should not include (0,1), (1,0)
Basically, we take the feature matrix, rotate it 45 degress,
and compute a "wavy" matrix for the features.
It's like the diagonal but it moves around to hopefully always include the true path.
"""
cdef np.ndarray[int, ndim=2] x_y_path_ = np.array(x_y_path).astype(np.int32)
assert (vecs0.shape[0] == norms0.shape[0])
assert (vecs1.shape[0] == norms1.shape[0])
assert (vecs0.shape[1] == norms0.shape[1])
assert (vecs1.shape[1] == norms1.shape[1])
# check how many overlaps vectors were passed in
num_overlaps_in_vecs0 = vecs0.shape[0]
num_overlaps_in_vecs1 = vecs1.shape[0]
# check how many overlaps were requested
# edge case: alignment_types could be empty
# In that case, we should just return insertions/deletions
# and max_x_overlap == max_y_overlap == 0
max_x_overlap = max([0] + [x for x, y in alignment_types]) # add [0] in case alignment_types is empty
max_y_overlap = max([0] + [y for x, y in alignment_types]) # add [0] in case alignment_types is empty
# note: alignment types are specified 1-based, but vectors are stored 0-based
if max_x_overlap > num_overlaps_in_vecs0:
raise Exception('%d x overlaps requrested (via alignment_types), but vecs0 only has %d' % (
max_x_overlap, num_overlaps_in_vecs0))
if max_y_overlap > num_overlaps_in_vecs1:
raise Exception('%d y overlaps requrested (via alignment_types), but vecs1 only has %d' % (
max_y_overlap, num_overlaps_in_vecs1))
# number of sentences in each document
cdef int xsize = vecs0.shape[1]
cdef int ysize = vecs1.shape[1]
# vector diminsions should match
assert (vecs0.shape[2] == vecs1.shape[2])
cdef np.ndarray[int, ndim=1] x_offsets, y_offsets
x_offsets, y_offsets = make_x_y_offsets(alignment_types)
# reserve outputs
a_len = x_y_path_.shape[0]
b_len = 2 * width_over2
cdef np.ndarray[float, ndim=3] a_b_feats = np.empty((len(alignment_types), a_len, b_len), dtype=np.float32)
cdef np.ndarray[int, ndim=1] b_offset = np.empty(a_len).astype(np.int32)
cdef int x, y, aa, bb, xx, yy, a_idx, b_idx, bb2, x_offset, y_offset, ii_align, x_offset_idx, y_offset_idx
cdef int vecsize = vecs0.shape[2]
cdef int num_alignments = x_offsets.shape[0]
cdef float sumx, feat
cdef float inf = np.inf
for ii in range(x_y_path_.shape[0]):
x = x_y_path_[ii, 0]
y = x_y_path_[ii, 1]
# convert xy to ab cords
aa = x + y
bb = y
a_idx = aa
b_offset[aa] = bb - width_over2
for b_idx, bb2 in enumerate(range(bb - width_over2, bb + width_over2)):
# convert ab to xy cords
xx = aa - bb2
yy = bb2
for ii_align in range(num_alignments):
x_offset = x_offsets[ii_align]
x_offset_idx = x_offset - 1 # overlaps start at 1, vectors stored 0-based
y_offset = y_offsets[ii_align]
y_offset_idx = y_offset - 1
if 0 <= xx < xsize and 0 <= yy < ysize:
sumx = 0.0
for jj in range(vecsize):
sumx += vecs0[x_offset_idx, xx, jj] * vecs1[y_offset_idx, yy, jj]
feat = 2.0 * x_offset * y_offset * (1.0 - sumx) / (
1e-6 + norms0[x_offset_idx, xx] + norms1[y_offset_idx, yy])
else:
feat = inf
a_b_feats[ii_align, a_idx, b_idx] = feat
return a_b_feats, b_offset
def sparse_dp(np.ndarray[float, ndim=3] a_b_costs,
np.ndarray[int, ndim=1] b_offset_in,
alignment_types,
double del_penalty,
int x_in_size,
int y_in_size):
"""
Do DP along a path, using features saved off along path.
x_offsets, y_offsets should not include (0,1), (1,0)
xsize, ysize refer to the costs a_b_csum, but in x/y space
As in the simpler full-DP case,
we compute cumulative costs and backpointers on notes,
and there are COSTS associated with moving between them.
This means the size of the notes +1,+1 larger (in x,y) than the COSTS.
So the size of a_b_csum, a_b_xp, a_b_yp are all one larger in x and y compared to the costs
In order to save memory (and time, vs a sparse matrix with hashes to look up values), let:
a = x + y
b = x - y
b_offsets tells us how far from the left edge the features are computed for.
basically it's like we are computing along the diagonal,
but we shift the diagonal around based on our belief
about where the alignments are.
b_offsets is used for both costs AND csum, backpointers, so it needs to be
+2 longer (it is in the a-direction) than the costs (in the a direction)
"""
cdef np.ndarray[int, ndim=1] x_offsets, y_offsets
x_offsets, y_offsets = make_x_y_offsets(alignment_types)
# make x/y offsets, including (0,1), (1,), i.e. including deletion and insertion
x_offsets = np.concatenate([x_offsets, np.array([0, 1], dtype=np.int32)])
y_offsets = np.concatenate([y_offsets, np.array([1, 0], dtype=np.int32)])
cdef int a_in_size = a_b_costs.shape[1]
cdef int b_in_size = a_b_costs.shape[2]
cdef int a_out_size = a_in_size + 2
cdef int b_out_size = b_in_size
cdef int x_out_size = x_in_size + 1
cdef int y_out_size = y_in_size + 1
# costs are the costs of going between nodes.
# in x,y for the nodes, we basically add a buffer
# at x=0 and y=0, and shift the cost by (x=+1,y=+1)
# In a,b space, this means adding two points (for the buffer)
# at the beginning, and shifting by (a=+0,b=+1) since
# a=x+y and b=y
# for the first two points, we can simply replicate the
# original b_offset, since it should be -width_over2
# i.e. b_offset_in[0] == -width_over2
extra_two_points = np.array([b_offset_in[0], b_offset_in[0]], dtype=np.int32)
cdef np.ndarray[int, ndim=1] b_offset_out = np.concatenate([extra_two_points, b_offset_in + 1])
# outputs
# For anything being used in accumulation, use float64
cdef np.ndarray[double, ndim=2] a_b_csum = np.zeros((a_in_size + 2, b_in_size),
dtype=np.float64) + np.inf # error cumulative sum
cdef np.ndarray[int, ndim=2] a_b_xp = np.zeros((a_in_size + 2, b_in_size), dtype=np.int32) - 2 # backpointer for x
cdef np.ndarray[int, ndim=2] a_b_yp = np.zeros((a_in_size + 2, b_in_size), dtype=np.int32) - 2 # backpointer for y
cdef int num_alignments = x_offsets.shape[0]
cdef double inf = np.inf
cdef int xx_out, yy_out, ii_align, x_offset, y_offset
cdef int aa_in_cost, bb_in_cost, aa_out, bb_out, aa_out_prev, bb_out_prev, xx_in_cost, yy_in_cost, xx_out_prev, yy_out_prev
cdef double alignment_cost, total_cost, prev_cost
# increasing in a is the same as going along diagonals in x/y, so DP order works
# (and any ordering is fine in b - nothing depends on values adjacent on diagonal in x/y)
for aa_out in range(a_in_size + 2):
for bb_out in range(b_in_size):
#xx_out, yy_out = ab2xy_w_offset(aa_out, bb_out, b_offset_out)
yy_out = bb_out + b_offset_out[aa_out]
xx_out = aa_out - yy_out
# edge case: all deletions in y-direction
if xx_out == 0 and 0 <= yy_out < y_out_size:
a_b_csum[aa_out, bb_out] = del_penalty * yy_out
a_b_xp[aa_out, bb_out] = 0
a_b_yp[aa_out, bb_out] = 1
# edge case: all deletions in x-direction
elif yy_out == 0 and 0 <= xx_out < x_out_size:
a_b_csum[aa_out, bb_out] = del_penalty * xx_out
a_b_xp[aa_out, bb_out] = 1
a_b_yp[aa_out, bb_out] = 0
else:
# initialize output to inf
a_b_csum[aa_out, bb_out] = inf
a_b_xp[aa_out, bb_out] = -42
a_b_yp[aa_out, bb_out] = -42
for ii_align in range(num_alignments):
x_offset = x_offsets[ii_align]
y_offset = y_offsets[ii_align]
# coords of location of alignment cost, in input x/y space
xx_in_cost = xx_out - 1 # features were front padded,
yy_in_cost = yy_out - 1 # so offset is always 1
# the coords of location of previous cumsum cost, in input x/y space
xx_out_prev = xx_out - x_offset
yy_out_prev = yy_out - y_offset
if 0 <= xx_in_cost < x_in_size and 0 <= yy_in_cost < y_in_size and 0 <= xx_out_prev < x_out_size and 0 <= yy_out_prev < y_out_size:
# convert x,y to a,b
aa_in_cost = xx_in_cost + yy_in_cost
bb_in_cost = yy_in_cost - b_offset_in[aa_in_cost]
aa_out_prev = xx_out_prev + yy_out_prev
bb_out_prev = yy_out_prev - b_offset_out[aa_out_prev]
if 0 <= aa_in_cost < a_in_size and 0 <= bb_in_cost < b_in_size and 0 <= aa_out_prev < a_out_size and 0 <= bb_out_prev < b_out_size:
if x_offset == 0 or y_offset == 0:
alignment_cost = del_penalty
else:
alignment_cost = a_b_costs[ii_align, aa_in_cost, bb_in_cost]
prev_cost = a_b_csum[aa_out_prev, bb_out_prev]
total_cost = prev_cost + alignment_cost
if total_cost < a_b_csum[aa_out, bb_out]:
a_b_csum[aa_out, bb_out] = total_cost
a_b_xp[aa_out, bb_out] = x_offset
a_b_yp[aa_out, bb_out] = y_offset
return a_b_csum, a_b_xp, a_b_yp, b_offset_out

665
ext-lib/vecalign/dp_utils.py Normal file
View File

@@ -0,0 +1,665 @@
"""
Copyright 2019 Brian Thompson
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
https://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License.
"""
import logging
import sys
from ast import literal_eval
from collections import OrderedDict
from math import ceil
from time import time
import numpy as np
import pyximport
pyximport.install(setup_args={'include_dirs':np.get_include()}, inplace=True, reload_support=True)
from dp_core import make_dense_costs, score_path, sparse_dp, make_sparse_costs, dense_dp
logger = logging.getLogger('vecalign') # set up in vecalign.py
def preprocess_line(line):
line = line.strip()
if len(line) == 0:
line = 'BLANK_LINE'
return line
def yield_overlaps(lines, num_overlaps):
lines = [preprocess_line(line) for line in lines]
for overlap in range(1, num_overlaps + 1):
for out_line in layer(lines, overlap):
# check must be here so all outputs are unique
out_line2 = out_line[:10000] # limit line so dont encode arbitrarily long sentences
yield out_line2
def read_in_embeddings(text_file, embed_file):
"""
Given a text file with candidate sentences and a corresponing embedding file,
make a maping from candidate sentence to embedding index,
and a numpy array of the embeddings
"""
sent2line = dict()
with open(text_file, 'rt', encoding="utf-8") as fin:
for ii, line in enumerate(fin):
if line.strip() in sent2line:
raise Exception('got multiple embeddings for the same line')
sent2line[line.strip()] = ii
line_embeddings = np.fromfile(embed_file, dtype=np.float32, count=-1)
if line_embeddings.size == 0:
raise Exception('Got empty embedding file')
laser_embedding_size = line_embeddings.size // len(sent2line) # currently hardcoded to 1024
if laser_embedding_size != 1024:
logger.warning('expected an embedding size of 1024, got %s', laser_embedding_size)
logger.info('laser_embedding_size determined to be %d', laser_embedding_size)
line_embeddings.resize(line_embeddings.shape[0] // laser_embedding_size, laser_embedding_size)
return sent2line, line_embeddings
def make_doc_embedding(sent2line, line_embeddings, lines, num_overlaps):
"""
lines: sentences in input document to embed
sent2line, line_embeddings: precomputed embeddings for lines (and overlaps of lines)
"""
lines = [preprocess_line(line) for line in lines]
vecsize = line_embeddings.shape[1]
vecs0 = np.empty((num_overlaps, len(lines), vecsize), dtype=np.float32)
for ii, overlap in enumerate(range(1, num_overlaps + 1)):
for jj, out_line in enumerate(layer(lines, overlap)):
try:
line_id = sent2line[out_line]
except KeyError:
logger.warning('Failed to find overlap=%d line "%s". Will use random vector.', overlap, out_line)
line_id = None
if line_id is not None:
vec = line_embeddings[line_id]
else:
vec = np.random.random(vecsize) - 0.5
vec = vec / np.linalg.norm(vec)
vecs0[ii, jj, :] = vec
return vecs0
def make_norm1(vecs0):
"""
make vectors norm==1 so that cosine distance can be computed via dot product
"""
for ii in range(vecs0.shape[0]):
for jj in range(vecs0.shape[1]):
norm = np.sqrt(np.square(vecs0[ii, jj, :]).sum())
vecs0[ii, jj, :] = vecs0[ii, jj, :] / (norm + 1e-5)
def layer(lines, num_overlaps, comb=' '):
"""
make front-padded overlapping sentences
"""
if num_overlaps < 1:
raise Exception('num_overlaps must be >= 1')
out = ['PAD', ] * min(num_overlaps - 1, len(lines))
for ii in range(len(lines) - num_overlaps + 1):
out.append(comb.join(lines[ii:ii + num_overlaps]))
return out
def read_alignments(fin):
alignments = []
with open(fin, 'rt', encoding="utf-8") as infile:
for line in infile:
fields = [x.strip() for x in line.split(':') if len(x.strip())]
if len(fields) < 2:
raise Exception('Got line "%s", which does not have at least two ":" separated fields' % line.strip())
try:
src = literal_eval(fields[0])
tgt = literal_eval(fields[1])
except:
raise Exception('Failed to parse line "%s"' % line.strip())
alignments.append((src, tgt))
# I know bluealign files have a few entries entries missing,
# but I don't fix them in order to be consistent previous reported scores
return alignments
def print_alignments(alignments, scores=None, file=sys.stdout):
if scores is not None:
for (x, y), s in zip(alignments, scores):
print('%s:%s:%.6f' % (x, y, s), file=file)
else:
for x, y in alignments:
print('%s:%s' % (x, y), file=file)
class DeletionKnob(object):
"""
A good deletion penalty is dependent on normalization, and probably language, domain, etc, etc
I want a way to control deletion penalty that generalizes well...
Sampling costs and use percentile seems to work fairly well.
"""
def __init__(self, samp, res_min, res_max):
self.res_min = res_min
self.res_max = res_max
if self.res_min >= self.res_max:
logger.warning('res_max <= res_min, increasing it')
self.res_max = self.res_min + 1e-4
num_bins = 1000
num_pts = 30
self.hist, self.bin_edges = np.histogram(samp, bins=num_bins,
range=[self.res_min, self.res_max],
density=True)
dx = self.bin_edges[1] - self.bin_edges[0]
self.cdf = np.cumsum(self.hist) * dx
interp_points = [(0, self.res_min), ]
for knob_val in np.linspace(0, 1, num_pts - 1)[1:-1]:
cdf_idx = np.searchsorted(self.cdf, knob_val)
cdf_val = self.res_min + cdf_idx / float(num_bins) * (self.res_max - self.res_min)
interp_points.append((knob_val, cdf_val))
interp_points.append((1, self.res_max))
self.x, self.y = zip(*interp_points)
def percentile_frac_to_del_penalty(self, knob_val):
del_pen = np.interp([knob_val], self.x, self.y)[0]
return del_pen
def make_alignment_types(max_alignment_size):
# return list of all (n,m) where n+m <= this
alignment_types = []
for x in range(1, max_alignment_size):
for y in range(1, max_alignment_size):
if x + y <= max_alignment_size:
alignment_types.append((x, y))
return alignment_types
def ab2xy_w_offset(aa, bb_idx, bb_offset):
bb_from_side = bb_idx + bb_offset[aa]
xx = aa - bb_from_side
yy = bb_from_side
return (xx, yy)
def xy2ab_w_offset(xx, yy, bb_offset):
aa = xx + yy
bb_from_side = yy
bb = bb_from_side - bb_offset[aa]
return aa, bb
def process_scores(scores, alignments):
# floating point sometimes gives negative numbers, which is a little unnerving ...
scores = np.clip(scores, a_min=0, a_max=None)
for ii, (x_algn, y_algn) in enumerate(alignments):
# deletion penalty is pretty arbitrary, just report 0
if len(x_algn) == 0 or len(y_algn) == 0:
scores[ii] = 0.0
# report sores un-normalized by alignment sizes
# (still normalized with random vectors, though)
else:
scores[ii] = scores[ii] / len(x_algn) / len(y_algn)
return scores
def sparse_traceback(a_b_csum, a_b_xp, a_b_yp, b_offset, xsize, ysize):
alignments = []
xx = xsize
yy = ysize
cum_costs = []
while True:
aa, bb = xy2ab_w_offset(xx, yy, b_offset)
cum_costs.append(a_b_csum[aa, bb])
xp = a_b_xp[aa, bb]
yp = a_b_yp[aa, bb]
if xx == yy == 0:
break
if xx < 0 or yy < 0:
raise Exception('traceback bug')
x_side = list(range(xx - xp, xx))
y_side = list(range(yy - yp, yy))
alignments.append((x_side, y_side))
xx = xx - xp
yy = yy - yp
alignments.reverse()
cum_costs.reverse()
costs = np.array(cum_costs[1:]) - np.array(cum_costs[:-1])
# "costs" are scaled by x_alignment_size * y_alignment_size
# and the cost of a deletion is del_penalty
# "scores": 0 for deletion/insertion,
# and cosine distance, *not* scaled
# by len(x_alignment)*len(y_alignment)
scores = process_scores(scores=costs, alignments=alignments)
return alignments, scores
def dense_traceback(x_y_tb):
xsize, ysize = x_y_tb.shape
xx = xsize - 1
yy = ysize - 1
alignments = []
while True:
if xx == yy == 0:
break
bp = x_y_tb[xx, yy]
if bp == 0:
xp, yp = 1, 1
alignments.append(([xx - 1], [yy - 1]))
elif bp == 1:
xp, yp = 0, 1
alignments.append(([], [yy - 1]))
elif bp == 2:
xp, yp = 1, 0
alignments.append(([xx - 1], []))
else:
raise Exception('got unknown value')
xx = xx - xp
yy = yy - yp
alignments.reverse()
return alignments
def append_slant(path, xwidth, ywidth):
"""
Append quantized approximation to a straight line
from current x,y to a point at (x+xwidth, y+ywidth)
"""
NN = xwidth + ywidth
xstart, ystart = path[-1]
for ii in range(1, NN + 1):
x = xstart + round(xwidth * ii / NN)
y = ystart + round(ywidth * ii / NN)
# In the case of ties we want them to round differently,
# so explicitly make sure we take a step of 1, not 0 or 2
lastx, lasty = path[-1]
delta = x + y - lastx - lasty
if delta == 1:
path.append((x, y))
elif delta == 2:
path.append((x - 1, y))
elif delta == 0:
path.append((x + 1, y))
def alignment_to_search_path(algn):
"""
Given an alignment, make searchpath.
Searchpath must step exactly one position in x XOR y at each time step.
In the case of a block of deletions, the order found by DP is not meaningful.
To make things consistent and to improve the probability of recovering
from search errors, we search an approximately straight line
through a block of deletions. We do the same through a many-many
alignment, even though we currently don't refine a many-many alignment...
"""
path = [(0, 0), ]
xdel, ydel = 0, 0
ydel = 0
for x, y in algn:
if len(x) and len(y):
append_slant(path, xdel, ydel)
xdel, ydel = 0, 0
append_slant(path, len(x), len(y))
elif len(x):
xdel += len(x)
elif len(y):
ydel += len(y)
append_slant(path, xdel, ydel)
return path
def extend_alignments(course_alignments, size0, size1):
"""
extend alignments to include new endpoints size0, size1
if alignments are larger than size0/size1, raise exception
"""
# could be a string of deletions or insertions at end, so cannot just grab last one
xmax = 0 # maximum x value in course_alignments
ymax = 0 # maximum y value in course_alignments
for x, y in course_alignments:
for xval in x:
xmax = max(xmax, xval)
for yval in y:
ymax = max(ymax, yval)
if xmax > size0 or ymax > size1:
raise Exception('asked to extend alignments but already bigger than requested')
# do not duplicate xmax/ymax, do include size0/size1
extra_x = list(range(xmax + 1, size0 + 1))
extra_y = list(range(ymax + 1, size1 + 1))
logger.debug('extending alignments in x by %d and y by %d', len(extra_x), len(extra_y))
if len(extra_x) == 0:
for yval in extra_y:
course_alignments.append(([], [yval]))
elif len(extra_y) == 0:
for xval in extra_x:
course_alignments.append(([xval], []))
else:
course_alignments.append((extra_x, extra_y))
def upsample_alignment(algn):
def upsample_one_alignment(xx):
return list(range(min(xx) * 2, (max(xx) + 1) * 2))
new_algn = []
for xx, yy in algn:
if len(xx) == 0:
for yyy in upsample_one_alignment(yy):
new_algn.append(([], [yyy]))
elif len(yy) == 0:
for xxx in upsample_one_alignment(xx):
new_algn.append(([xxx], []))
else:
new_algn.append((upsample_one_alignment(xx), upsample_one_alignment(yy)))
return new_algn
def make_del_knob(e_laser,
f_laser,
e_laser_norms,
f_laser_norms,
sample_size):
e_size = e_laser.shape[0]
f_size = f_laser.shape[0]
if e_size > 0 and f_size > 0 and sample_size > 0:
if e_size * f_size < sample_size:
# dont sample, just compute full matrix
sample_size = e_size * f_size
x_idxs = np.zeros(sample_size, dtype=np.int32)
y_idxs = np.zeros(sample_size, dtype=np.int32)
c = 0
for ii in range(e_size):
for jj in range(f_size):
x_idxs[c] = ii
y_idxs[c] = jj
c += 1
else:
# get random samples
x_idxs = np.random.choice(range(e_size), size=sample_size, replace=True).astype(np.int32)
y_idxs = np.random.choice(range(f_size), size=sample_size, replace=True).astype(np.int32)
# output
random_scores = np.empty(sample_size, dtype=np.float32)
score_path(x_idxs, y_idxs,
e_laser_norms, f_laser_norms,
e_laser, f_laser,
random_scores, )
min_score = 0
max_score = max(random_scores) # could bump this up... but its probably fine
else:
# Not much we can do here...
random_scores = np.array([0.0, 0.5, 1.0]) # ???
min_score = 0
max_score = 1 # ????
del_knob = DeletionKnob(random_scores, min_score, max_score)
return del_knob
def compute_norms(vecs0, vecs1, num_samples, overlaps_to_use=None):
# overlaps_to_use = 10 # 10 matches before
overlaps1, size1, dim = vecs1.shape
overlaps0, size0, dim0 = vecs0.shape
assert (dim == dim0)
if overlaps_to_use is not None:
if overlaps_to_use > overlaps1:
raise Exception('Cannot use more overlaps than provided. You may want to re-run make_verlaps.py with a larger -n value')
else:
overlaps_to_use = overlaps1
samps_per_overlap = ceil(num_samples / overlaps_to_use)
if size1 and samps_per_overlap:
# sample other size (from all overlaps) to compre to this side
vecs1_rand_sample = np.empty((samps_per_overlap * overlaps_to_use, dim), dtype=np.float32)
for overlap_ii in range(overlaps_to_use):
idxs = np.random.choice(range(size1), size=samps_per_overlap, replace=True)
random_vecs = vecs1[overlap_ii, idxs, :]
vecs1_rand_sample[overlap_ii * samps_per_overlap:(overlap_ii + 1) * samps_per_overlap, :] = random_vecs
norms0 = np.empty((overlaps0, size0), dtype=np.float32)
for overlap_ii in range(overlaps0):
e_laser = vecs0[overlap_ii, :, :]
sim = np.matmul(e_laser, vecs1_rand_sample.T)
norms0[overlap_ii, :] = 1.0 - sim.mean(axis=1)
else: # no samples, no normalization
norms0 = np.ones((overlaps0, size0)).astype(np.float32)
return norms0
def downsample_vectors(vecs1):
a, b, c = vecs1.shape
half = np.empty((a, b // 2, c), dtype=np.float32)
for ii in range(a):
# average consecutive vectors
for jj in range(0, b - b % 2, 2):
v1 = vecs1[ii, jj, :]
v2 = vecs1[ii, jj + 1, :]
half[ii, jj // 2, :] = v1 + v2
# compute mean for all vectors
mean = np.mean(half[ii, :, :], axis=0)
for jj in range(0, b - b % 2, 2):
# remove mean
half[ii, jj // 2, :] = half[ii, jj // 2, :] - mean
# make vectors norm==1 so dot product is cosine distance
make_norm1(half)
return half
def vecalign(vecs0,
vecs1,
final_alignment_types,
del_percentile_frac,
width_over2,
max_size_full_dp,
costs_sample_size,
num_samps_for_norm,
norms0=None,
norms1=None):
if width_over2 < 3:
logger.warning('width_over2 was set to %d, which does not make sense. increasing to 3.', width_over2)
width_over2 = 3
# make sure input embeddings are norm==1
make_norm1(vecs0)
make_norm1(vecs1)
# save off runtime stats for summary
runtimes = OrderedDict()
# Determine stack depth
s0, s1 = vecs0.shape[1], vecs1.shape[1]
max_depth = 0
while s0 * s1 > max_size_full_dp ** 2:
max_depth += 1
s0 = s0 // 2
s1 = s1 // 2
# init recursion stack
# depth is 0-based (full size is 0, 1 is half, 2 is quarter, etc)
stack = {0: {'v0': vecs0, 'v1': vecs1}}
# downsample sentence vectors
t0 = time()
for depth in range(1, max_depth + 1):
stack[depth] = {'v0': downsample_vectors(stack[depth - 1]['v0']),
'v1': downsample_vectors(stack[depth - 1]['v1'])}
runtimes['Downsample embeddings'] = time() - t0
# compute norms for all depths, add sizes, add alignment types
t0 = time()
for depth in stack:
stack[depth]['size0'] = stack[depth]['v0'].shape[1]
stack[depth]['size1'] = stack[depth]['v1'].shape[1]
stack[depth]['alignment_types'] = final_alignment_types if depth == 0 else [(1, 1)]
if depth == 0 and norms0 is not None:
if norms0.shape != vecs0.shape[:2]:
print('norms0.shape:', norms0.shape)
print('vecs0.shape[:2]:', vecs0.shape[:2])
raise Exception('norms0 wrong shape')
stack[depth]['n0'] = norms0
else:
stack[depth]['n0'] = compute_norms(stack[depth]['v0'], stack[depth]['v1'], num_samps_for_norm)
if depth == 0 and norms1 is not None:
if norms1.shape != vecs1.shape[:2]:
print('norms1.shape:', norms1.shape)
print('vecs1.shape[:2]:', vecs1.shape[:2])
raise Exception('norms1 wrong shape')
stack[depth]['n1'] = norms1
else:
stack[depth]['n1'] = compute_norms(stack[depth]['v1'], stack[depth]['v0'], num_samps_for_norm)
runtimes['Normalize embeddings'] = time() - t0
# Compute deletion penalty for all depths
t0 = time()
for depth in stack:
stack[depth]['del_knob'] = make_del_knob(e_laser=stack[depth]['v0'][0, :, :],
f_laser=stack[depth]['v1'][0, :, :],
e_laser_norms=stack[depth]['n0'][0, :],
f_laser_norms=stack[depth]['n1'][0, :],
sample_size=costs_sample_size)
stack[depth]['del_penalty'] = stack[depth]['del_knob'].percentile_frac_to_del_penalty(del_percentile_frac)
logger.debug('del_penalty at depth %d: %f', depth, stack[depth]['del_penalty'])
runtimes['Compute deletion penalties'] = time() - t0
tt = time() - t0
logger.debug('%d x %d full DP make features: %.6fs (%.3e per dot product)',
stack[max_depth]['size0'], stack[max_depth]['size1'], tt,
tt / (stack[max_depth]['size0'] + 1e-6) / (stack[max_depth]['size1'] + 1e-6))
# full DP at maximum recursion depth
t0 = time()
stack[max_depth]['costs_1to1'] = make_dense_costs(stack[max_depth]['v0'],
stack[max_depth]['v1'],
stack[max_depth]['n0'],
stack[max_depth]['n1'])
runtimes['Full DP make features'] = time() - t0
t0 = time()
_, stack[max_depth]['x_y_tb'] = dense_dp(stack[max_depth]['costs_1to1'], stack[max_depth]['del_penalty'])
stack[max_depth]['alignments'] = dense_traceback(stack[max_depth]['x_y_tb'])
runtimes['Full DP'] = time() - t0
# upsample the path up to the top resolution
compute_costs_times = []
dp_times = []
upsample_depths = [0, ] if max_depth == 0 else list(reversed(range(0, max_depth)))
for depth in upsample_depths:
if max_depth > 0: # upsample previoius alignment to current resolution
course_alignments = upsample_alignment(stack[depth + 1]['alignments'])
# features may have been truncated when downsampleing, so alignment may need extended
extend_alignments(course_alignments, stack[depth]['size0'], stack[depth]['size1']) # in-place
else: # We did a full size 1-1 search, so search same size with more alignment types
course_alignments = stack[0]['alignments']
# convert couse alignments to a searchpath
stack[depth]['searchpath'] = alignment_to_search_path(course_alignments)
# compute ccosts for sparse DP
t0 = time()
stack[depth]['a_b_costs'], stack[depth]['b_offset'] = make_sparse_costs(stack[depth]['v0'], stack[depth]['v1'],
stack[depth]['n0'], stack[depth]['n1'],
stack[depth]['searchpath'],
stack[depth]['alignment_types'],
width_over2)
tt = time() - t0
num_dot_products = len(stack[depth]['b_offset']) * len(stack[depth]['alignment_types']) * width_over2 * 2
logger.debug('%d x %d sparse DP (%d alignment types, %d window) make features: %.6fs (%.3e per dot product)',
stack[max_depth]['size0'], stack[max_depth]['size1'],
len(stack[depth]['alignment_types']), width_over2 * 2,
tt, tt / (num_dot_products + 1e6))
compute_costs_times.append(time() - t0)
t0 = time()
# perform sparse DP
stack[depth]['a_b_csum'], stack[depth]['a_b_xp'], stack[depth]['a_b_yp'], \
stack[depth]['new_b_offset'] = sparse_dp(stack[depth]['a_b_costs'], stack[depth]['b_offset'],
stack[depth]['alignment_types'], stack[depth]['del_penalty'],
stack[depth]['size0'], stack[depth]['size1'])
# performace traceback to get alignments and alignment scores
# for debugging, avoid overwriting stack[depth]['alignments']
akey = 'final_alignments' if depth == 0 else 'alignments'
stack[depth][akey], stack[depth]['alignment_scores'] = sparse_traceback(stack[depth]['a_b_csum'],
stack[depth]['a_b_xp'],
stack[depth]['a_b_yp'],
stack[depth]['new_b_offset'],
stack[depth]['size0'],
stack[depth]['size1'])
dp_times.append(time() - t0)
runtimes['Upsample DP compute costs'] = sum(compute_costs_times[:-1])
runtimes['Upsample DP'] = sum(dp_times[:-1])
runtimes['Final DP compute costs'] = compute_costs_times[-1]
runtimes['Final DP'] = dp_times[-1]
# log time stats
max_key_str_len = max([len(key) for key in runtimes])
for key in runtimes:
if runtimes[key] > 5e-5:
logger.info(key + ' took ' + '.' * (max_key_str_len + 5 - len(key)) + ('%.4fs' % runtimes[key]).rjust(7))
return stack

BIN
ext-lib/vecalign/media/dynamic_programing_approximation.gif Normal file
View File

Binary file not shown.
Side by Side

After

Width:  |  Height:  |  Size: 1.0 MiB

BIN
ext-lib/vecalign/media/multilingual_sentence_embedding.png Normal file
View File

Binary file not shown.
Side by Side

After

Width:  |  Height:  |  Size: 56 KiB

61
ext-lib/vecalign/overlap.py Normal file
View File

@@ -0,0 +1,61 @@
#!/usr/bin/env python3
"""
Copyright 2019 Brian Thompson
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
https://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License.
"""
import argparse
from dp_utils import yield_overlaps
def go(output_file, input_files, num_overlaps):
output = set()
for fin in input_files:
lines = open(fin, 'rt', encoding="utf-8").readlines()
for out_line in yield_overlaps(lines, num_overlaps):
output.add(out_line)
# for reproducibility
output = list(output)
output.sort()
with open(output_file, 'wt', encoding="utf-8") as fout:
for line in output:
fout.write(line + '\n')
def _main():
parser = argparse.ArgumentParser('Create text file containing overlapping sentences.',
formatter_class=argparse.ArgumentDefaultsHelpFormatter)
parser.add_argument('-i', '--inputs', type=str, nargs='+',
help='input text file(s).')
parser.add_argument('-o', '--output', type=str,
help='output text file containing overlapping sentneces')
parser.add_argument('-n', '--num_overlaps', type=int, default=4,
help='Maximum number of allowed overlaps.')
args = parser.parse_args()
go(output_file=args.output,
num_overlaps=args.num_overlaps,
input_files=args.inputs)
if __name__ == '__main__':
_main()

170
ext-lib/vecalign/score.py Normal file
View File

@@ -0,0 +1,170 @@
#!/usr/bin/env python3
"""
Copyright 2019 Brian Thompson
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
https://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License.
"""
import argparse
import sys
from collections import defaultdict
import numpy as np
from dp_utils import read_alignments
"""
Faster implementation of lax and strict precision and recall, based on
https://www.aclweb.org/anthology/W11-4624/.
"""
def _precision(goldalign, testalign):
"""
Computes tpstrict, fpstrict, tplax, fplax for gold/test alignments
"""
tpstrict = 0 # true positive strict counter
tplax = 0 # true positive lax counter
fpstrict = 0 # false positive strict counter
fplax = 0 # false positive lax counter
# convert to sets, remove alignments empty on both sides
testalign = set([(tuple(x), tuple(y)) for x, y in testalign if len(x) or len(y)])
goldalign = set([(tuple(x), tuple(y)) for x, y in goldalign if len(x) or len(y)])
# mappings from source test sentence idxs to
# target gold sentence idxs for which the source test sentence
# was found in corresponding source gold alignment
src_id_to_gold_tgt_ids = defaultdict(set)
for gold_src, gold_tgt in goldalign:
for gold_src_id in gold_src:
for gold_tgt_id in gold_tgt:
src_id_to_gold_tgt_ids[gold_src_id].add(gold_tgt_id)
for (test_src, test_target) in testalign:
if (test_src, test_target) == ((), ()):
continue
if (test_src, test_target) in goldalign:
# strict match
tpstrict += 1
tplax += 1
else:
# For anything with partial gold/test overlap on the source,
# see if there is also partial overlap on the gold/test target
# If so, its a lax match
target_ids = set()
for src_test_id in test_src:
for tgt_id in src_id_to_gold_tgt_ids[src_test_id]:
target_ids.add(tgt_id)
if set(test_target).intersection(target_ids):
fpstrict += 1
tplax += 1
else:
fpstrict += 1
fplax += 1
return np.array([tpstrict, fpstrict, tplax, fplax], dtype=np.int32)
def score_multiple(gold_list, test_list, value_for_div_by_0=0.0):
# accumulate counts for all gold/test files
pcounts = np.array([0, 0, 0, 0], dtype=np.int32)
rcounts = np.array([0, 0, 0, 0], dtype=np.int32)
for goldalign, testalign in zip(gold_list, test_list):
pcounts += _precision(goldalign=goldalign, testalign=testalign)
# recall is precision with no insertion/deletion and swap args
test_no_del = [(x, y) for x, y in testalign if len(x) and len(y)]
gold_no_del = [(x, y) for x, y in goldalign if len(x) and len(y)]
rcounts += _precision(goldalign=test_no_del, testalign=gold_no_del)
# Compute results
# pcounts: tpstrict,fnstrict,tplax,fnlax
# rcounts: tpstrict,fpstrict,tplax,fplax
if pcounts[0] + pcounts[1] == 0:
pstrict = value_for_div_by_0
else:
pstrict = pcounts[0] / float(pcounts[0] + pcounts[1])
if pcounts[2] + pcounts[3] == 0:
plax = value_for_div_by_0
else:
plax = pcounts[2] / float(pcounts[2] + pcounts[3])
if rcounts[0] + rcounts[1] == 0:
rstrict = value_for_div_by_0
else:
rstrict = rcounts[0] / float(rcounts[0] + rcounts[1])
if rcounts[2] + rcounts[3] == 0:
rlax = value_for_div_by_0
else:
rlax = rcounts[2] / float(rcounts[2] + rcounts[3])
if (pstrict + rstrict) == 0:
fstrict = value_for_div_by_0
else:
fstrict = 2 * (pstrict * rstrict) / (pstrict + rstrict)
if (plax + rlax) == 0:
flax = value_for_div_by_0
else:
flax = 2 * (plax * rlax) / (plax + rlax)
result = dict(recall_strict=rstrict,
recall_lax=rlax,
precision_strict=pstrict,
precision_lax=plax,
f1_strict=fstrict,
f1_lax=flax)
return result
def log_final_scores(res):
print(' ---------------------------------', file=sys.stderr)
print('| | Strict | Lax |', file=sys.stderr)
print('| Precision | {precision_strict:.3f} | {precision_lax:.3f} |'.format(**res), file=sys.stderr)
print('| Recall | {recall_strict:.3f} | {recall_lax:.3f} |'.format(**res), file=sys.stderr)
print('| F1 | {f1_strict:.3f} | {f1_lax:.3f} |'.format(**res), file=sys.stderr)
print(' ---------------------------------', file=sys.stderr)
def main():
parser = argparse.ArgumentParser(
'Compute strict/lax precision and recall for one or more pairs of gold/test alignments',
formatter_class=argparse.ArgumentDefaultsHelpFormatter)
parser.add_argument('-t', '--test', type=str, nargs='+', required=True,
help='one or more test alignment files')
parser.add_argument('-g', '--gold', type=str, nargs='+', required=True,
help='one or more gold alignment files')
args = parser.parse_args()
if len(args.test) != len(args.gold):
raise Exception('number of gold/test files must be the same')
gold_list = [read_alignments(x) for x in args.gold]
test_list = [read_alignments(x) for x in args.test]
res = score_multiple(gold_list=gold_list, test_list=test_list)
log_final_scores(res)
if __name__ == '__main__':
main()

162
ext-lib/vecalign/vecalign.py Normal file
View File

@@ -0,0 +1,162 @@
#!/usr/bin/env python3
"""
Copyright 2019 Brian Thompson
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
https://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License.
"""
"""
Usage:
python ext-lib/vecalign/vecalign.py \
-s data/mac/dev/zh \
-t data/mac/dev/en \
-o data/mac/dev/auto \
-m data/mac/dev/meta_data.tsv \
--src_embed data/mac/dev/zh/overlap data/mac/dev/zh/overlap.emb \
--tgt_embed data/mac/dev/en/overlap data/mac/dev/en/overlap.emb \
-a 8 -v
"""
import os
import time
import argparse
import shutil
import logging
import pickle
from math import ceil
from random import seed as seed
import numpy as np
logger = logging.getLogger('vecalign')
logger.setLevel(logging.WARNING)
logFormatter = logging.Formatter("%(asctime)s %(levelname)-5.5s %(message)s")
consoleHandler = logging.StreamHandler()
consoleHandler.setFormatter(logFormatter)
logger.addHandler(consoleHandler)
from dp_utils import make_alignment_types, read_alignments, read_in_embeddings, make_doc_embedding, vecalign
def main():
# make runs consistent
seed(42)
np.random.seed(42)
parser = argparse.ArgumentParser('Sentence alignment using Vecalign')
parser.add_argument('-s', '--src', type=str, required=True,
help='preprocessed source file to align')
parser.add_argument('-t', '--tgt', type=str, required=True,
help='preprocessed target file to align')
parser.add_argument('-o', '--out', type=str, required=True,
help='Output directory.')
parser.add_argument('-m', '--meta', type=str, required=True,
help='Metadata file.')
parser.add_argument('--src_embed', type=str, nargs=2, required=True,
help='Source embeddings. Requires two arguments: first is a text file, sencond is a binary embeddings file. ')
parser.add_argument('--tgt_embed', type=str, nargs=2, required=True,
help='Target embeddings. Requires two arguments: first is a text file, sencond is a binary embeddings file. ')
parser.add_argument('-a', '--alignment_max_size', type=int, default=5,
help='Searches for alignments up to size N-M, where N+M <= this value. Note that the the embeddings must support the requested number of overlaps')
parser.add_argument('-d', '--del_percentile_frac', type=float, default=0.2,
help='Deletion penalty is set to this percentile (as a fraction) of the cost matrix distribution. Should be between 0 and 1.')
parser.add_argument('-v', '--verbose', help='sets consle to logging.DEBUG instead of logging.WARN',
action='store_true')
parser.add_argument('--max_size_full_dp', type=int, default=300,
help='Maximum size N for which is is acceptable to run full N^2 dynamic programming.')
parser.add_argument('--costs_sample_size', type=int, default=20000,
help='Sample size to estimate costs distribution, used to set deletion penalty in conjunction with deletion_percentile.')
parser.add_argument('--num_samps_for_norm', type=int, default=100,
help='Number of samples used for normalizing embeddings')
parser.add_argument('--search_buffer_size', type=int, default=5,
help='Width (one side) of search buffer. Larger values makes search more likely to recover from errors but increases runtime.')
args = parser.parse_args()
if args.verbose:
import logging
logger.setLevel(logging.INFO)
if args.alignment_max_size < 2:
logger.warning('Alignment_max_size < 2. Increasing to 2 so that 1-1 alignments will be considered')
args.alignment_max_size = 2
src_sent2line, src_line_embeddings = read_in_embeddings(args.src_embed[0], args.src_embed[1])
tgt_sent2line, tgt_line_embeddings = read_in_embeddings(args.tgt_embed[0], args.tgt_embed[1])
width_over2 = ceil(args.alignment_max_size / 2.0) + args.search_buffer_size
make_dir(args.out)
jobs = create_jobs(args.meta, args.src, args.tgt, args.out)
for rec in jobs:
src_file, tgt_file, align_file = rec.split("\t")
logger.info('Aligning src="%s" to tgt="%s"', src_file, tgt_file)
src_lines = open(src_file, 'rt', encoding="utf-8").readlines()
vecs0 = make_doc_embedding(src_sent2line, src_line_embeddings, src_lines, args.alignment_max_size)
tgt_lines = open(tgt_file, 'rt', encoding="utf-8").readlines()
vecs1 = make_doc_embedding(tgt_sent2line, tgt_line_embeddings, tgt_lines, args.alignment_max_size)
final_alignment_types = make_alignment_types(args.alignment_max_size)
logger.debug('Considering alignment types %s', final_alignment_types)
stack = vecalign(vecs0=vecs0,
vecs1=vecs1,
final_alignment_types=final_alignment_types,
del_percentile_frac=args.del_percentile_frac,
width_over2=width_over2,
max_size_full_dp=args.max_size_full_dp,
costs_sample_size=args.costs_sample_size,
num_samps_for_norm=args.num_samps_for_norm)
# write final alignments
print_alignments(stack[0]['final_alignments'], align_file)
def create_jobs(meta, src, tgt, out):
jobs = []
fns = get_fns(meta)
for file in fns:
src_path = os.path.abspath(os.path.join(src, file))
tgt_path = os.path.abspath(os.path.join(tgt, file))
out_path = os.path.abspath(os.path.join(out, file + '.align'))
jobs.append('\t'.join([src_path, tgt_path, out_path]))
return jobs
def get_fns(meta):
fns = []
with open(meta, 'rt', encoding='utf-8') as f:
next(f) # skip header
for line in f:
recs = line.strip().split('\t')
fns.append(recs[0])
return fns
def print_alignments(alignments, out):
with open(out, 'wt', encoding='utf-8') as f:
for x, y in alignments:
f.write("{}:{}\n".format(x, y))
def make_dir(path):
if os.path.isdir(path):
shutil.rmtree(path)
os.makedirs(path, exist_ok=True)
if __name__ == '__main__':
t_0 = time.time()
main()
print("It takes {} seconds to aligent all the sentences.".format(time.time() - t_0))