Section: New Results

Advances in statistical parsing

Participants : Marie Candito, Benoît Crabbé, Djamé Seddah, Enrique Henestroza Anguiano.

Improving statistical dependency parsing

Alpage has provided state-of-the art results for French statistical Parsing, adapting existing techniques for French, a richer morphological language than English, either for constituency parsing or dependency parsing. The Bonsai tool (see section  5.4 ) is available, that gathers preprocessing tools and models for dependency parsing French. We have innovated in the tuning of tagsets and the handling of unknown words. In the last years, Alpage has then contributed on four main points:

• conversion of the French Treebank [59] used as constituency training data into dependencies [72] , the resulting treebank being used by several teams for dependency parsing;

• an original method to reduce lexical data sparseness and include coverage and robustness by replacing tokens by unsupervised word clusters or morphological clusters [69] , [121] , [73] ; all of our morphological clustering approaches were integrated into our parsing chains; data driven lemmatization required the adaptation of a state-of-the-art part-of-speech tagger and lemmatizer (Morfette [77] ) based on a data-driven joint model benefiting of the inclusion of external lexica such as the Lefff [121] .

• a parser-agnostic postprocessing step, developed this year, which uses specialized models for dependency parse correction [30] : dependencies in an input parse tree are revised by selecting, for a given dependent, the best governor from within a small set of candidates, using a discriminative linear ranking model that includes a rich feature set that encodes syntactic structure in the input parse tree; the parse correction framework can correct attachments using either a generic model or specialized models tailored to difficult attachment types like coordination and pp-attachment; our experiments have shown that parse correction, combining a generic model with specialized models for difficult attachment types, can successfully improve the quality of predicted parse trees output by several representative state-of-the-art dependency parsers for French.

• an adaptation of the above-mentioned technique of word clustering to the problem of adapting statistical parsers to different text domains [25] . We show that in order to parse texts from a different domain than the one a statistical parser is trained on (namely to parse target domain text using a parser trained on indomain treebank), word clusters computed over a bridge corpus that couples indomain an target domain raw texts do improve parsing performance on target domain, without degrade performance on indomain texts (contrary to previous domain adaptation techniques). To evaluate these experiments, we use as target domain biomedical texts. We have supervised the manual syntactic annotation of a test corpus from the biomedical domain (European Public Assessment Reports concerning the marketing authorization of medicinal products).

Besides this line of work, it should be noted that two parsing models built around Stochastic Tree Insertion Grammars are currently under investigation: experiments have been conducted on Spinal TIGs [122] . Moreover, we are still improving the TIG-based dependency parser MICA, developed in collaboration with University of Marseilles, Columbia university and AT&T [61] (see section  5.5 ).

Functional labelling

Alpage worked towards the improvement of a functional labeller to be used as a post-parsing tool on an unfolded parse forest (as outputted e.g. by the Berkeley parser in the Bonsai architecture) using CRF models of various orders thereby extending the previous maximum entropy labeller designed in the team. The use of CRFs for modelling triggered a collaboration with Isabelle Tellier and JP Prost (LIFO, Orleans). The labeller implementation has been considerably improved and the accurracy of the labeller has improved as well on correct treebank trees. However we found out that the feature engineering work outweights the formal improvements since we were able to show that the use of higher order graphical models were not contributing significantly to improve an unstructured model. Our modest gains come mostly from feature engineering. Moreover we notice that combined with a constituent parser the labeller does not improve at all on constituent parsing output. The reason being that our current architecture for the Bonsai parser is sequential (which is unsatisfactory). Following experiments on n-best parsing outputs, we observe that the labeller can drastically improve on better parses where its input is indeed correct. This suggest investigating formulating constituent parsing and functional labelling as a joint task requiring to address serious efficiency issues. We intend to tackle the two drawbacks of our current architecture (sequential process, parse forest unfolding) by formulating constituent parsing as a joint task with functional labelling in the next few months.

Parsing spontaneous oral text

Alpage also got involved in parsing spontaneous oral text taken from ESTER 3 data (with overlaps) generated in the ANR ETAPE project in collaboration with A. Abeillé (LLF) with the aim of preannotating a seed for a future treebank of oral French which would considerably support work in experimental linguistics led in the Labex. He has also a collaboration set up with A. Abeillé, C. Gardent and C. Cerisara for ensuring interoperability accross ongoing efforts for producing oral treebanks for French. The way to carry out the task was by using a form of preprocessing of oral text to simulate a written entry to the Bonsai parser trained on written text. In the next few months we intend to test semi-supervised learning techniques to speed up the annotation process made by the LLF lab.