Adapting the RASP System for the CoNLL07 Domain-Adaptation Task

Proceedings of the CoNLL Shared Task Session of EMNLP-CoNLL 2007, pp. 1170–1174, Prague, June 2007. c⃝2007 Association for Computational Linguistics Adapting the RASP System for the CoNLL07 Domain-Adaptation Task Rebecca Watson and Ted Briscoe Computer Laboratory University of Cambridge FirstName.LastName@cl.cam.ac.uk Abstract We describe our submission to the domain adaptation track of the CoNLL07 shared task in the open class for systems using ex- ternal resources. Our main finding was that it was very difficult to map from the annota- tion scheme used to prepare training and de- velopment data to one that could be used to effectively train and adapt the RASP system unlexicalized parse ranking model. Never- theless, we were able to demonstrate a sig- nificant improvement in performance utiliz- ing bootstrapping over the PBIOTB data. 1 Introduction The CoNLL07 domain adaptation task was created to explore how a parser trained in one domain might be adapted to a new one. The training data were drawn from the PTB (Marcus et al., 1993) rean- notated with dependency relations (Johansson and Nugues, 2007, hereafter DRs). The test data were taken from a corpus of biomedical articles (Kulick et al., 2004) and the CHILDES database (Brown, 1973; MacWhinney, 2000) also reannotated with DRs (see Nivre et al., 2007) for further details of the task, annotation format, and evaluation scheme. The development data consisted of a small amount of annotated and unannotated biomedical and con- versational data. The RASP system (Briscoe et al., 2006) utilizes a manually-developed grammar and outputs gram- matical bilexical dependency relations (see Briscoe, 2006 for a detailed description, hereafter GRs). Wat- son et al. (2007) describe a semi-supervised, boot- strapping approach to training the parser which uti- lizes unlabelled partially-bracketed input with re- spect to the system derivations. For the domain adaptation task we retrained RASP by mapping our GR scheme to the DR scheme and annotation for- mat, and used this mapping to select a derivation to train the unlexicalized parse ranking model from the annotated PTB training data. We also performed similar partially-supervised bootstrapping over the 200 annotated biomedical sentences in the develop- ment data. We then tried unsupervised bootstrap- ping from the unannotated development data based on these initial models. As the parser requires input to consist of a se- quence of one of 150 CLAWS PoS tags, we also uti- lize a first-order HMM PoS tagger which has been trained on manually-annotated data from the LOB, BNC and Susanne Corpora (see Briscoe, 2006 for further details). Accordingly, we submitted our re- sults in the open class. 2 Training and Adaptation The RASP parser is a generalized LR parser which builds a non-deterministic generalized LALR(1) parse table from the grammar (Tomita, 1987). A context-free ‘backbone’ is automatically derived from a unification grammar. The residue of fea- tures not incorporated into the backbone are unified on each reduce action and if unification fails the as- sociated derivation paths also fail. The parser cre- ates a packed parse forest represented as a graph- structured stack. Inui et …

Content truncated to 500 words