Because Syntax Does Matter: Improving Predicate-Argument Structures Parsing with Syntactic Features

Human Language Technologies: The 2015 Annual Conference of the North American Chapter of the ACL, pages 64–74, Denver, Colorado, May 31 – June 5, 2015. c⃝2015 Association for Computational Linguistics Because Syntax Does Matter: Improving Predicate-Argument Structures Parsing with Syntactic Features Corentin Ribeyre⋆ ◦ Eric Villemonte de la Clergerie⋆ Djamé Seddah⋆ ⋄ ⋆Alpage, INRIA ◦Univ Paris Diderot, Sorbonne Paris Cité ⋄ Université Paris Sorbonne firstname.lastname@inria.fr Abstract Parsing full-fledged predicate-argument struc- tures in a deep syntax framework requires graphs to be predicted. Using the DeepBank (Flickinger et al., 2012) and the Predicate- Argument Structure treebank (Miyao and Tsu- jii, 2005) as a test field, we show how transition-based parsers, extended to handle connected graphs, benefit from the use of topologically different syntactic features such as dependencies, tree fragments, spines or syntactic paths, bringing a much needed con- text to the parsing models, improving notably over long distance dependencies and elided coordinate structures. By confirming this pos- itive impact on an accurate 2nd-order graph- based parser (Martins and Almeida, 2014), we establish a new state-of-the-art on these data sets. 1 Introduction For the majority of the state-of-the-art parsers that routinely reach ninety percent performance plateau in capturing tree structures, the question of what next crucially arises. Indeed, it has long been thought that the bottleneck preventing the advent of accu- rate syntax-to-semantic interfaces lies in the qual- ity of the preceding phase of analysis: the better the parse, the better the output. The truth is that most of the structures used to train current parsing mod- els are degraded versions of a more informative data set: the Wall Street journal section of the Penn tree- bank (PTB, (Marcus et al., 1993)) which is often stripped of its richer set of annotations (i.e. traces and functional labels are removed), while, for rea- sons of efficiency and availability, projective depen- dency trees are often given preference over richer graph structures (Nivre and Nilsson, 2005; Sagae and Tsujii, 2008). This led to the emergence of sur- face syntax-based parsers (Charniak, 2000; Nivre, 2003; Petrov et al., 2006) whose output cannot by themselves be used to extract full-fledged predicate- argument structures. For example, control verb con- structions, it-cleft structures, argument sharing in el- lipsis coordination, etc. are among the phenomena requiring a graph to be properly accounted for. The dichotomy between what can usually be parsed with high accuracy and what lies in the deeper syntac- tic description has initiated a line of research de- voted to closing the gap between surface syntax and richer structures. For most of the previous decade, the term deep syntax was used for rich parsing mod- els built upon enriched versions of a constituency treebank, either with added HPSG or LFG annota- tion or CCG (almost) full rewrites (Miyao and Tsu- jii, 2005; Cahill et al., 2004; Hockenmaier, 2003). Its use now spreads by misnomer to models that pro- vide more abstract structures, capable of generaliz- ing classical functional labels to more semantic (in a logical view) arguments, potentially capable …

Content truncated to 500 words