Hard Time Parsing Questions: Building a QuestionBank for French

Hard Time Parsing Questions: Building a QuestionBank for French Djamé Seddah⋆ ⋄ Marie Candito⋆ ◦ ⋆Alpage, INRIA ⋄ Université Paris-Sorbonne ◦Université Paris Diderot djame.seddah@paris-sorbonne.fr marie.candito@linguist.univ-paris-diderot.fr Abstract We present the French Question Bank, a treebank of 2600 questions. We show that classical parsing model performance drop while the inclusion of this data set is highly beneficial without harming the parsing of non-question data. when facing out-of-domain data with strong structural divergences. Two thirds being aligned with the English QuestionBank (Judge et al., 2006) and being freely available, this treebank will prove useful to build robust NLP systems. Keywords: Treebanking; Parsing; Question-phrase 1. Introduction One of the ever-recurring issues in statistical parsing is the matter of out-of-domain parsing. Namely how to make a parser trained on, by definition, a narrow domain able to cope with any kind of text. The range of possible issues can be characterized on a 3-axis graph where each axis de- notes the divergence compared to an in-domain treebank, well edited, treebank: (i) a lexical divergence axis, (ii) a noisy-ness axis and (iii) a syntactic divergence axis. In this work, we focus on the syntactic divergences underly- ing the question genre. To do so, we present the French Question Bank (FQB), a French treebank of syntactically- annotated questions1, and we investigate the performance of in-domain trained parsers on this data set, showing a clear loss of performance brought by structural divergences at the functional level. When added to the training set, large improvements are shown, demonstrating the usefulness of this new data set. 2. French questions typology Our motivation was to improve statistical parsing perfor- mance on questions, which are crucial for e.g. QA and yet difficult to (statistically) parse due to the often non- canonical structure and word order they exhibit. More pre- cisely, we can distinguish roughly the questions with an extracted (i.e. fronted) phrase from the in situ questions, which exhibit the canonical word order. In situ questions can be split into: (i) those containing a wh-phrase, namely a constituent with embedded inter- rogative determiner, adjective, pronoun or adverb2, but ap- pearing in canonical position (for instance : Paul a mangé quel dessert? lit. ’Paul has eaten which dessert?’ (Which dessert did Paul eat?), or (ii) yes/no questions, for which the interrogative status is either marked by prosody / ques- tion mark only (Paul a déjà mangé ? lit. ’Paul has already eaten?’, or using a nominative clitic after the inflected verb. The clitic is either redundant (clitic doubling) with a non- anaphoric pre-verbal subject (Paul a-t-il déjà mangé? ’Paul 1This is to our knowledge the first non-English QuestionBank. 2The main ones are qui (who), que (what), quel (which), quoi (what[-hum]), quand (when), où (where), comment (how), ... Wh- words in French are sometimes called “mot-qu”, as the “wh” French counterpart is “qu”. has-CL-NOM-3rd-sg already eaten?’ (Has Paul already eaten?)) or it is anaphoric and plays the role of the subject (A-t-il déjà mangé? ’Has-CL-NOM-3rd-sg already eaten?’ (Has he already eaten?)). Extracted wh- phrase Questions …

Content truncated to 500 words