Latest commit 9866b64
Mar 25, 2013
|Failed to load latest commit information.|
// README // // 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. // // Copyright 2004-2013 Brian Roark // Author: firstname.lastname@example.org (Brian Roark) Incremental Top-Down Parser README The incremental top-down parser is a statistical syntactic parser that processes input strings from left-to-right, producing partial derivations in a top-down manner, using beam search as detailed in Roark (2001) and Roark (2004). It can output parser state statistics of utility for psycholinguistic studies, as detailed in Roark et al. (2009). Note that this code is research code, provided to the community without promise of additional technical support. Many researchers have used this code in their research, and were able to successfully compile and use the code on linux and Mac OS X systems. If you follow the instructions below, you should be able to use it as well, but if you encounter problems you are pretty much on your own. In addition to the parser and the model training code, a model trained on sections 2-21 of the Penn WSJ treebank has been included. For those who want to train their own models, there are some scripts for converting Penn Treebank files from the 'pretty-print' representation provided by the LDC to the format expected by the model training routine: one parse per line, with no function tags or empty nodes, etc. You are free to change your non-terminal labels (e.g., changing VBs to AUXs, etc.) but the code assumes that POS-tags and non-POS-tag non-terminals are disjoint. ====== INSTALLATION ======= To compile, you will need to change the install directory in Makedefs, then cd to the bin directory and call make all. Linux requires -lm flag, Mac OSX doesn't. Copy Makefile.MAC to Makefile for Mac OSX; copy Makefile.linux to Makefile for linux systems. ====== USAGE ======= Parsing: To get information on usage, type the following on the command line: bin/tdparse -? The following is standard usage for parsing a file of strings 'mystrings.txt': bin/tdparse -v -F parse.output parse.model.wsj.slc.p05 mystrings.txt The strings file 'mystrings.txt' should have one string per line (see the file data/test.test.txt). The parser output comes with three numbers before the parse: cand tot-cands -logCondProb (TOP... This is for distinguishing between candidates in k-best parsing. To evaluate with just one candidate: sed 's/.*(TOP/(TOP/g' parse.output >parse.best Then: evalb -p data/evalb/COLLINS.prm mytrees.gold.txt parse.best >parse.eval Training models: To train a model, you must first prepare a training treebank. Scripts have been provided to assist in converting Penn Treebank style parses from their raw format as distributed by the LDC to the format expected by the model training code. Briefly, that code expects: 1) 1 parse per line in the text file; 2) no empty nodes; 3) a single TOP category, with only unary productions to the root non-terminal of the tree; 4) disjoint POS-tag and non-POS-tag non-terminal sets. Our scripts make use of tsurgeon available here: http://nlp.stanford.edu/software/tregex.shtml Once that has been installed, modify the data/tbnorm/tbnorm.sh script to include the path to the tsurgeon directory. Then, cd to data/tbnorm and: ./tbnorm.sh raw.treebank.file >normalized.treebank.file This script does not guarantee condition 4 above, that the POS-tags are disjoint from the other non-terminals. The tbtest.sh script tests that this is the case: ./tbtest.sh normalized.treebank.file This will report non-terminals used as both POS-tags and non-POS-tag non-terminals in the given treebank file. See data/test.train.txt to see a small, fake treebank in the required format. To get information on usage, type the following on the command line: bin/tdptrain -? The following is standard usage for training a model from a given treebank: bin/tdptrain -t0.05 -p -s"(SINV(S" -v -F parse.model.slc.p05 -m data/tree.functs.sfile normalized.treebank.file Please see the usage from the command line for further options. ==================== new verbose output: If you run with the -p command line switch, it outputs word-by-word scores that are available from standard running of the parser, including the surprisal measures and other internal state statistics. There are column headers, and I'll explain each of them in turn: pfix header In this column, we get some info on the particular row. For each word w, this will read 'pfix#w' where # can range over three symbols: '-' are tokens which 'belong' to the previous token, such as punctuation or contractions or end-of-string; '+' are closed class lexical items; and ':' are anything else. non-lexical and closed class words can are identified by including the following switch when parsing: -c data/closed.class.words and that file can be modified as you like. In that file, 0 following the word means punct. or contraction; 1 following the word means closed class. prefix This column gives prefix probability, sum of probs over the beam. srprsl nolex lex These three columns give total surprisal, syntactic surprisal, and lexical surprisal, as defined in the paper. ambig This column gives the entropy over the current beam, essentially a measure of ambiguity. If p is the conditional probability of each analysis in the normalized beam, then this is -\sum p log p over the beam. open This column gives the weighted average open brackets over the beam. (Easy to derive from the parser, not of any particular utility so far.) rernk toprr These two columns examine what happens to the previously top-ranked parses. I'll explain this informally by abusing notation. Let T_i be the best scoring parse after word i, and let T^+_i be the set of parses after word i+1 that are extensions of parse T_i. Then rernk is p(T^+_i)/p(T_i) where the p in this case is normalized over the beam (hence ratio can be greater than 1). A score of 1 means that these analyses occupy about the same amount of conditional probability mass in the beam after one step; a score greater than 1 means they are now occupying more of the conditional probability mass; and less than 1 means they occupy less. A score of zero means it has fallen out of the beam. The second column here just looks at the top ranked parse at time i+1. If it is an extension of the top ranked parse at time i, then the ratio of conditional prob is provided; otherwise zero. stps Finally, we look at the weighted average (over the normalized beam) of derivation steps from the previous word. Part II: measures that cannot be derived from standard operation of the parser. These include the entropy measures, which are calculated by extending the current parse to all possible following words (hence running with this switch is slllllloooooowwwwww). To get the entropy scores, run with the -a switch, and it will provide additional scores related to entropy. afix header just like prefix header, with closed class coding, etc. entropy noLexH LexH total entropy, Syntactic entropy and lexical entropy, as described in the paper. rank log10(rank) These two columns show the rank of the actual next word within the vocabulary, in terms of the conditional prob of the word given the left context. Since we are calculating the probability of every word in the lexicon, we can straightforwardly rank and report where the actual word falls in the rank (and the log). log(p*/p) This score compares the probability of the top ranked word in the lexicon versus the actual next word. Thus if rank is 1, this score will be zero, otherwise greater than zero. One last thing. There are scores attached to each output tree (by default the 1-best, but use -k to retrieve more), and those scores are: candidate rank total number of candidates in list tree score (without norm constant) tree negative log prob purely syntactic contribution to that total neg log prob lexical contribution to the neg log prob (from POS -> word) =========== CITING ================ The basic citation for the parsing approach is: Brian Roark. 2001. Probabilistic top-down parsing and language modeling. In Computational Linguistics, 27(2), pages 249-276. If using the parser to derive parser state statistics for psycholinguistic modeling, it's also a good idea to cite: Brian Roark, Asaf Bachrach, Carlos Cardenas and Christophe Pallier. 2009. Deriving lexical and syntactic expectation-based measures for psycholinguistic modeling via incremental top-down parsing. In Proceedings of the Conference on Empirical Methods in Natural Language Processing (EMNLP), pp. 324-333. =========== REFERENCES ================ Brian Roark. 2011. Expected surprisal and entropy. Technical Report #CSLU-11-004, Center for Spoken Language Processing, Oregon Health & Science University. Brian Roark, Asaf Bachrach, Carlos Cardenas and Christophe Pallier. 2009. Deriving lexical and syntactic expectation-based measures for psycholinguistic modeling via incremental top-down parsing. In Proceedings of the Conference on Empirical Methods in Natural Language Processing (EMNLP), pp. 324-333. Michael Collins and Brian Roark. 2004. Incremental Parsing with the Perceptron Algorithm. In Proceedings of the 42nd Annual Meeting of the Association for Computational Linguistics (ACL), pages 111-118. Brian Roark. 2004. Robust garden path parsing. Natural Language Engineering, 10(1), pages 1-24. Brian Roark. 2001. Robust Probabilistic Predictive Syntactic Processing: Motivations, Models, and Applications. Ph.D. Thesis, Department of Cognitive and Linguistic Sciences, Brown University. Brian Roark. 2001. Probabilistic top-down parsing and language modeling. In Computational Linguistics, 27(2), pages 249-276. Mark Johnson and Brian Roark. 2000. Compact non-left-recursive grammars using the selective left-corner transform and factoring. In Proceedings of the 18th International Conference on Computational Linguistics (COLING), pages 355-361. Brian Roark and Mark Johnson. 1999. Efficient probabilistic top-down and left-corner parsing. In Proceedings of the 37th Annual Meeting of the Association for Computational Linguistics, pages 421-428.