PASCAL Challenge on Grammar Induction (Shared Task)
This shared task aims to foster continuing research in grammar induction and part-of-speech induction, while also opening up the problem to more ambitious settings, including a wider variety of languages, removing the reliance on gold standard parts-of-speech and, critically, providing a thorough evaluation including a task-based evaluation.
The shared task will evaluate dependency grammar induction algorithms, evaluating the quality of structures induced from natural language text. In contrast with the defacto standard experimental setup, which starts with gold standard part-of-speech tags, we will encourage competitors to submit systems which are completely unsupervised. The evaluation will consider the standard dependency tree based measures (directed, undirected edge accuracy, bracketing accuracy etc) as well as measures over the predicted parts of speech. Our aim is to allow a wide range of different approaches, and for this reason we will accept submissions which predict just the dependency trees for gold PoS, just the PoS, or both jointly.
While our focus is on unsupervised approaches, we recognise that there has been considerable related research using semi-supervised learning, domain adaption, cross-lingual projection and other partially supervised methods for building syntactic models. We will support these kinds of systems, but require the participants to declare which external resources they have used. When presenting the results, we will split them into two sets: purely unsupervised approaches and those that have some form of external supervision.
Notice: The Test Data submission date is coming Friday, April 13th.
Please send your test predictions by that date, by posting to the WILS Google group, with the predictions in an archive attached to the message. Please also state which part of the challenge you're participating in: POS induction, dependency tree induction or both jointly. Lastly, we'd like one or two paragraphs describing your algorithm, including an outline of your algorithm, its data requirements and any novel features of your approach.
Shared Task Dates
Training Data Released
Test Data Submission
Gold Test Data Release
Evaluation results released
Shared Task System Description Submission
Jun 4-6, 2012
Jun 7-8, 2012
Gold Part-of-Speech tags - Dependency structures should be predicted using the gold tagset (universal or not?)
Induced Part-of-Speech tags - Predicting dependency structures and/or Part-of-Speech tags directly from the text. The number of induced tags can be chosen by each participant.
Open resources - Other resources can be used, such as parallel data.
If you plan to participate, please join the google group. We'll post announcements and any questions about the data, evaluation etc to this list. Nb: You must contact the organisers to get access to the LDC data - we will send you the licence form which you should then fill in and fax to the LDC).
The data that we are providing was collated from existing treebanks in a variety of different languages, domains and linguistic formalisms. Specifically, we are using the following:
English Penn Treebank v3 WSJ
Czech Prague Dependency Treebank v2
Arabic Prague Arabic Dependency Treebank v1
Please download the files using the above download links. These corpora have open licence agreements for research purposes, and can be freely downloaded. Please see the README file for each corpus, which includes details of the licensing.
The first three corpora listed above are licensed under the LDC, who have agreed to allow competitors access to the data for the purpose of the competition. Participants will need to sign the special license agreement and send this to the LDC in order to gain access to these corpora (English PTB, Czech PDT, Arabic PADT).
New 17 April The annotated test data is now available. These can now be downloaded for the languages with open licences. Please contact the LDC for the test data for English PTB, Czech and Arabic.
Note that some of these corpora have been used in previous evaluations, namely the shared tasks at CONLL-X and CONLL 2007. In most cases our data is not identical, as we have updated these corpora to include larger amounts of data and changes to the treebanks that have occurred since the CONLL competitions. In addition, our data format is slightly different in order to include universal PoS tags.
Our multi-lingual setup is designed to allow competitors to develop cross-lingual approaches for transferring syntactic knowledge between languages. To support these techniques, we will evaluate competing systems against the fine tag-set, coarse tag-set and its reduction into Petrov et al.'s universal tag-set.
For the English PTB, we will compile multiple annotations for the same sentences such that the effect of the choice of linguistic formalism or annotation procedure can be offset in the evaluation. This is a long-standing issue in parsing where many researchers evaluate only against the Penn Treebank, a setting which does not reflect competing widely supported linguistic theories. Overall this test set will form a significant resource for the evaluation of parsers and grammar induction algorithms, and help to reduce the field's continuing reliance on the Penn Treebank.
All data files are encoded in UTF-8, and largely follow the file format from the CONLL-X/2007 shared tasks. Each sentence is represented as a series of lines, with one token per line, and sentences are split by blank lines. Each token is represented as a tab-separated list of fields:
- word number, starting from 1
- coarse part-of-speech (optional)
- fine part-of-speech
- universal part-of-speech
- lexical features (optional)
- index of head word
- type of dependency relation linking head to current word
Missing values are denoted by an underscore (_), and not all corpora include values for the optional fields. Each different corpus uses different annotation methods for tokenization, lemmatisation, lexical features, part-of-speech and dependency edge labels. Please see the README file in each corpus for descriptions of these annotations. The universal parts-of-speech, however, use the same tag-set across the different corpora.
For each corpus, we will distribute a large training set, a small development set with all fields and a test set. For the Gold Part-of-Speech stream of the competition, fields 8-9 will be omitted (replaced with an underscore) from the training and test sets, and only provided for the test set at the end of the competition. For the Induced Part-of-Speech stream, fields 4-9 will be omitted for the training and test sets. Note that as our task is induction, participants are encouraged to pool all the data together for the purpose of training their unsupervised models (i.e., use the union of training, development and testing sets for training their models).
note: To clarify: the results for sentences shorter than 10 words are reported by filtering on length before removing punctuation. The dependency evaluation script had a switch to filter on length, but this was not used in the shared task. The switch has been removed from the scripts, and a simple filtering script has been added. We apologise for any confusion caused.
New scripts are now available for
Running some baseline systems for POS induction, DMV dependency induction and a pipeline to do both tasks. There are also some simple non-trained dependency baselines included, e.g., right-branching.
Evaluating the induced POS and dependency trees. This includes many of the standard clustering metrics and common dependency evaluation metrics (un/directed accuracy, NED [Schwartz et al., 2011]).
Please note that python 2.7 is required for running the scripts.
See above for the evaluation scripts, covering
dependencies against a gold standard. This will measure the directed and undirected unlabelled attachment score, as well as Neutral Edge Detection [Schwartz et al., 2011] both for all sentences and sentences shorter than 10 words.
parts-of-speech against the gold standard, using each of the fine tags, coarse tags and universal tags. Clustering based approaches will be supported using the standard metrics for evaluating cluster identifiers, e.g., many-to-1, 1-1, VI etc.
Suggestions are welcome for other evaluation methods.
New 20 April The results are now available. See the following pages:
ResultsPos for POS induction results
ResultsDep for dependency induction results
ResultsPosDep for joint results
We suggest a Baseline system that can be used as a starting point for the shared task. The Baseline is an implementation of the Dependency Model with Valence (DMV)[Klein, 2004] as described in the paper [Gillenwater, 2010]. See the baseline download above under tools.
[Klein, 2004] - Corpus-based induction of syntactic structure: Models of dependency and constituency, D. Klein and C. Manning. In Proc. ACL, 2004.
[Gillenwater, 2010] - Posterior Sparsity in Dependency Grammar Induction, J. Gillenwater, K. Ganchev, J. Graca, F. Pereira, and B. Taskar. Journal of Machine Learning Research (JMLR).
[Schwartz et al., 2011] - Neutralizing Linguistically Problematic Annotations in Unsupervised Dependency Parsing Evaluation, Roy Schwartz, Omri Abend, Roi Reichart and Ari Rappoport. ACL 2011