Python scripts preprocessing Penn Treebank and Chinese Treebank
Switch branches/tags
Nothing to show
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Permalink
Failed to load latest commit information.
.gitignore
LICENSE Initial commit Nov 4, 2017
README.md CTB supports segmentation and pos-tagging Mar 4, 2018
ctb.py Documentation Jul 10, 2018
ctb8.py Support for ctb8.0 Jul 10, 2018
ptb.py
tb_to_stanford.py Support for converting Dependency Parsing corpus in .conllx format Mar 2, 2018
utility.py Support for ctb8.0 Jul 10, 2018

README.md

TreebankPreprocessing

Python scripts preprocessing Penn Treebank (PTB) and Chinese Treebank 5.1 (CTB). They can convert treebanks to:

Corpus Format Description
constituency parse tree .txt one line for one sentence
dependency parse tree .conllx Basic Stanford Dependencies (SD)
word segmentation corpus .tsv first column for characters, second column for BMES tags, sentences separated by a blank line
part-of-speech tagging corpus .tsv first column for words, second column for tags, sentences separated by a blank line

When designing a tagger or parser, preprocessing treebanks is a troublesome problem. We need to:

  • Split dataset into train/dev/test, following conventional splits.
  • Remove xml tags inside CTB.
  • Combine the multiline bracketed files into one file, one line for one sentence.

I wondered why there were no open-source tools handling these tedious works. Finally I decide to write one myself. Hopefully it will save you some time.

Required software

  • Python3
  • NLTK
  • Optional stanford-parser for converting to dependency parse trees

Overview

What kind of task can we perform on treebanks?

Chinese Word Segmentation

For CTB, segmentation corpus are split as per Jiang et al. (2009):

  • CTB Training: 001–270, 400–1151. Development: 301–325. Test: 271-300.

Part-of-Speech Tagging

  • PTB Training: 0-18. Development: 19-21. Test: 22-24. As per Collins (2002) and Choi (2016).
  • CTB The same with Chinese Word Segmentation.

Phrase Structure Parsing

These scripts can also convert treebanks into the conventional data setup from Chen and Manning (2014), Dyer et al. (2015). The detailed splits are:

  • PTB Training: 02-21. Development: 22. Test: 23.
  • CTB Training: 001–815, 1001–1136. Development: 886–931, 1148–1151. Test: 816–885, 1137–1147.

Dependency Parsing

You will need Stanford Parser for converting phrase structure trees to dependency parse trees. Please download the Stanford Parser Version 3.3.0 and place them in this folder:

TreebankPreprocessing
├── ...
├── stanford-parser-3.3.0-models.jar
└── stanford-parser.jar

OK, let's do it on the fly.

PTB

1. Import PTB into NLTK

Bracketed files parsing relies on NLTK. Please follow NLTK instruction, put BROWN and WSJ into nltk_data/corpora/ptb, e.g.

ptb
├── BROWN
└── WSJ

2. Run ptb.py

This script does all the work for you, only requires a path to store output.

$ python3 ptb.py --help 
usage: ptb.py [-h] --output OUTPUT [--task TASK]

Combine Penn Treebank WSJ MRG files into train/dev/test set

optional arguments:
  -h, --help       show this help message and exit
  --output OUTPUT  The folder where to store the output train/dev/test files
  --task TASK      Which task (par, pos)? Use par for phrase structure
                   parsing, pos for part-of-speech tagging
  • You will get 3 .txt files corresponding to train/dev/test set.
  • If you want part-of-speech tagging corpora, simply append --task pos. This time, you get 3 .tsv files.
  • .txt files can be converted to .conllx files by tb_to_stanford.py:
$ python3 tb_to_stanford.py --help
usage: tb_to_stanford.py [-h] --input INPUT --lang LANG --output OUTPUT

Convert combined Penn Treebank files (.txt) to Stanford Dependency format
(.conllx)

optional arguments:
  -h, --help       show this help message and exit
  --input INPUT    The folder containing train.txt/dev.txt/test.txt in
                   bracketed format
  --lang LANG      Which language? Use en for English, cn for Chinese
  --output OUTPUT  The folder where to store the output
                   train.conllx/dev.conllx/test.conllx in Stanford Dependency
                   format

CTB

The CTB is a little messy, it contains extra xml tags in every gold tree, and is not natively supported by NLTK. You need to specify the CTB root path (the folder containing index.html).

$ python3 ctb.py --help           
usage: ctb.py [-h] --ctb CTB --output OUTPUT [--task TASK]

Combine Chinese Treebank 5.1 fid files into train/dev/test set

optional arguments:
  -h, --help       show this help message and exit
  --ctb CTB        The root path to Chinese Treebank 5.1
  --output OUTPUT  The folder where to store the output
                   train.txt/dev.txt/test.txt
  --task TASK      Which task (seg, pos, par)? Use seg for word segmentation,
                   pos for part-of-speech tagging, par for phrase structure
                   parsing
  • Tagging and dependency parsing corpora can be obtained similar to PTB.

Then you can start your research, enjoy it!