uparse

Dependency Parsing Experiments using Makefile

Makefile included in uparse directory allows you to perform several dependency parsing experiments using MaltOptimizer and Maltparser

To perform experiments, edit Makefile to set CONLL_DIR location to your CoNLL format pentreebank corpus location. Directory you will provide should include a separate directory for each section of the corpus (00-24)

Now you can start experiments. Here are a few make calls for the experiments

• make all will execute a default experiment by performing parameter optimisation using section 22, train an optimised dependency parser using sections 2-21 and test the parser using section 23
• make all TASK_IDENTIFIER=ws.50 TRANSFORMOPT="--tagmode tagfile --tagfile ../data/upos/ws.50.gz" will first replace all tags/cpostags in given corpus by tags given with --tagfile option and will initiate make all

All experiments will create a result.<TASK_IDENTIFIER>.tar.gz file into uparse directory. This file contains

• model.eval including LA, LAS, UAS performance metrics
• model.out including parser output for section 23
• phase3_optFile.txt including Maltparser options
• addMergPOSTAGI0FORMStack0.xml including features to be used by Maltparser.

conll.py

usage: conll.py [-h] [--file FILE] [--directory DIRECTORY]
                [--extension EXTENSION] [--section SECTION]
                [--tagmode {nochange,tagfile,onetagperword,remove,random}]
                [--tagfile TAGFILE] [--ambigious]
                [--formmode {nochange,formfile,remove}] [--formfile FORMFILE]
                [--subsmode {nochange,best}] [--subsfile SUBSFILE]

CoNLL file transformer

optional arguments:
  -h, --help            show this help message and exit
  --file FILE           CoNLL files to read
  --directory DIRECTORY
                        CoNLL directory to read
  --extension EXTENSION
                        CoNLL file extension to read
  --section SECTION     WSJ sections to be filtered out
  --tagmode {nochange,tagfile,onetagperword,remove,random}
                        Tag manipulation operation on corpus
  --tagfile TAGFILE     Tag file to be used. Only valid when used with
                        --tagmode tagfile|random
  --ambigious           Threat tags as ambigious by not using a dictionary
  --formmode {nochange,formfile,remove}
                        Form manipulation operation on corpus
  --formfile FORMFILE   Form file to be used. Only valid when used with
                        --formmode formfile
  --subsmode {nochange,best}
                        Form manipulation operation on corpus
  --subsfile SUBSFILE   Substitution file to be used. Only valid when used
                        with --subsmode formfile

filter.py

$ python2.7 filter.py -h
usage: filter.py [-h] [--wsj10 | --wsj20 | --wsj40 | --wsj]
                 inputwildcard output

Cleanup & Filter CoNLL corpus

positional arguments:
  inputwildcard  File wildcard showing CoNLL corpus file(s) including one or
                 more dependency graphs
  output         Target CoNLL corpus file used to store clean and filtered
                 CoNLL corpus

optional arguments:
  -h, --help     show this help message and exit
  --wsj10        Cleaned sentences of maximum length 10
  --wsj20        Cleaned sentences of maximum length 20
  --wsj40        Cleaned sentences of maximum length 40
  --wsj          Cleaned sentences of full corpus

eval.py

$ python2.7 eval.py -h
usage: eval.py [-h] [--ignoreroot] [--minlength MINLENGTH] goldfile modelfile

Evaluate two parsings

positional arguments:
  goldfile              Source CoNLL corpus file including gold dependency
                        graphs
  modelfile             Model CoNLL corpus file including model dependency
                        grapgs

optional arguments:
  -h, --help            show this help message and exit
  --ignoreroot
  --minlength MINLENGTH
                        Minimum sentence length to be considered in evaluation

parser.py

$ python2.7 parser.py -h
usage: parser.py [-h] [--rhead | --lhead] input output

General some popular baseline parsings for given corpus files

positional arguments:
  input       Source CoNLL corpus file including gold dependency graphs
  output      Model CoNLL corpus file including model dependency grapgs

optional arguments:
  -h, --help  show this help message and exit
  --rhead     Right head parsing
  --lhead     Left head parsing

genconll.py

$ python2.7 genconll.py -h
usage: genconll.py [-h] [--parallel parallel]
                   input output [sections [sections ...]]

Generate CoNLL format by reading penntreebank trees

positional arguments:
  input                Treebank directory containing sections of penntree
                       corpus
  output               Target CoNLL directory
  sections             Section filter for generation

optional arguments:
  -h, --help           show this help message and exit
  --parallel parallel  Number of parallel slaves to perform conversion

###PCFG

./treetransform.py --brush --cnf ./data/treebank/treebank.mrg > ./data/json/_penntreebank.json

./retag.py --noextendedtag ./data/json/_penntreebank.json > ./data/json/penntreebank.json

./retag.py --tagfile ./data/upos/ws.100.gz ./data/nlp/treebank/treebank-2.0/json/wsj/wsj.json > ./data/nlp/treebank/treebank-2.0/json/wsj/wsj.ws.100.json

./split.py --tagged --maxlength 40 0.80 ./data/nlp/treebank/treebank-2.0/json/wsj/wsj.json

./count_cfg_freq.py ./data/nlp/treebank/treebank-2.0/json/wsj/wsj.train > penntreebank.counts

./pcfg_parse.py --parallel 8 penntreebank.counts ./data/nlp/treebank/treebank-2.0/json/wsj/wsj.dev > ./data/nlp/treebank/treebank-2.0/json/wsj/wsj.p1.out

./eval_parser.py ./data/nlp/treebank/treebank-2.0/json/wsj/wsj.key data/nlp/treebank/treebank-2.0/json/wsj/wsj.p1.out