CCG Parsing: 2015

This is the software used for the following publication:

Weakly-Supervised Grammar-Informed Bayesian CCG Parser Learning
Dan Garrette, Chris Dyer, Jason Baldridge, and Noah A. Smith
In Proceedings of AAAI 2015

A Supertag-Context Model for Weakly-Supervised CCG Parser Learning
Dan Garrette, Chris Dyer, Jason Baldridge, and Noah A. Smith
In Proceedings of CoNLL 2015

Getting the code

$ get clone git@github.com:dhgarrette/2015-ccg-parsing.git
$ cd 2015-ccg-parsing

Data setup

Put the English, Chinese, and Italian data into the following directories:

 data/ccgbank
 data/ccgbank-chinese
 data/ccgbank-italian

The files should be arranged as follows:

$ ls data/ccgbank/AUTO
00	02	04	06	08	10	12	14	16	18	20	22	24
01	03	05	07	09	11	13	15	17	19	21	23
$ ls data/ccgbank-chinese/AUTO
00	02	04	06	08	10	20	22	24	26	28	30
01	03	05	07	09	11	21	23	25	27	29	31
$ ls data/ccgbank-italian/pro
civil_law.pro.txt	jrc_acquis.pro.txt	newspaper.pro.txt    

Running the code

First, compile the code and generate the run script:

$ ./compile

Then run:

$ target/start dhg.ccg.run.Parse2015Run [options]

Options

• --model: The model to use. Options: (no default)
  • pcfg: for the pcfg model (see AAAI-2015 paper)
  • scg: for the supertag-context model (see CoNLL-2015 paper)
• --learning: The learning algorithm to use to train the model. Options: {mcmc}. Default: mcmc.
• --additional-rules: Additional CCG rules to be allowed by the parser (comma-separated). Example: FC,BX,FC2,BX2. Default x (meaning no additional rules).
• --lang: The language of the CCGBank to use. Options: {en, ch, it}. Default: en.
• --max-sent-len: The maximum sentence length allowed (filter all sentence longer than this). Options: an integer or all for no limit. Default: all.
• --td-tok: The maximum number of tokens to be read when building the tag dictionary. Options: an integer (following an integer with k will expand to 000; e.g. 10k becomes 10000) or all for no limit. Default: all.
• --train-sent: The maximum number of sentences to be used for training. Options: an integer or all for no limit. Default: all.
• --test-sent: The maximum number of sentences to be used for testing. Options: an integer or all for no limit. Default: all.
• --sampling-iterations: The number of MCMC sampling iterations to run. Default: 500.
• --burnin-iterations: The number of MCMC burn-in iterations to run. Default: 0.
• --alpha-root: See paper for details. Default: 1.0.
• --alpha-biny: See paper for details. Default: 100.0.
• --alpha-unry: See paper for details. Default: 100.0.
• --alpha-term: See paper for details. Default: 10000.0.
• --alpha-prod: See paper for details. Default: 100.0.
• --alpha-cntx: See paper for details. Only relevant for --model scg. Default: 1000.0.
• --root-init: Root parameter initializer. Options:
  • uniform
  • catprior: use the grammar-defined category prior.
  • tdecatprior: use the grammar-defined category prior, with atomic category probabilities estimated using the tag dictionary and raw data. DEFAULT.
• --nt-prod-init: Nonterminal production parameter initializer (for both binary and unary). Options:
  • uniform.
  • catprior: use the grammar-defined category prior.
  • tdecatprior: use the grammar-defined category prior, with atomic category probabilities estimated using the tag dictionary and raw data. DEFAULT.
• --term-prod-init: Terminal production parameter initializer (for both binary and unary). Options:
  • uniform.
  • tdentry: Use the tag dictionary and raw data to estimate terminal (word) probabilities for each supertag. DEFAULT.
• --tr-init: Context production parameter initializer (for both left and right contexts). Only relevant for --model scg. Options:
  • uniform,
  • tdentry (use the tag dictionary and raw data to estimate transition probabilities),
  • combine-uniform (use CCG supertag combinability mixed with uniform),
  • combine-tdentry (use CCG supertag combinability mixed with tdentry). DEFAULT.
• --pterm: See paper for details. Default: 0.7.
• --pmod: See paper for details. Default: 0.1.
• --pfwd: See paper for details. Default: 0.5.
• --comb-tr-mass: Amount of probability mass devoted to "combinable" contexts (called sigma (σ) in the CoNLL-2015 paper). Only relevant for --model scg. Default: 0.85.
• --td-cutoff: Exclude tag dictionary entries that occur with less than this proportion in the TD-training corpus. Default: 0.0.
• --max-accept-tries: Number of samples drawn for each sentence in each iteration. Only relevant for --model scg. Default: 1.
• --output-file: A file where the parsed trees of the test sentences should be written. Default: do not write out trees.
• --train-termdel: Allow terminal deletion from a training sentence when a parse is not found. Choices {false,true}. Default: false.
• --test-termdel: Allow terminal deletion from a test sentence when a parse is not found. Choices {false,true}. Default: false.
• --max-train-tok: The maximum number of tokens to be read for the training data. Options: an integer (following an integer with k will expand to 000; e.g. 10k becomes 10000) or all for no limit. Default: all.
• --mcmc-output-count-file: File where additional data should be written. Default: do not write out this information.