Encoding Generalized Quantifiers in Dependency-based Compositional Semantics

PACLIC 28 !585 Encoding Generalized Quantifiers in Dependency-based Compositional Semantics Yubing Dong⇤ Department of Computer Science University of Southern California yubing.dong@usc.edu Ran Tian Graduate School of Information Sciences Tohoku University tianran@ecei.tohoku.ac.jp Yusuke Miyao National Institute of Informatics, Japan yusuke@nii.ac.jp Abstract For textual entailment recognition systems, it is often important to correctly handle Gen- eralized quantifiers (GQ). In this paper, we explore ways of encoding GQs in a recent framework of Dependency-based Composi- tional Semantics, especially aiming to cor- rectly handle linguistic knowledge like hy- ponymy when GQs are involved. We use both the selection operator mechanism and a new relation extension to implement some major properties of GQs, reducing 69% errors of a previous system, and a further error analy- sis suggests extensions towards more power- ful logical systems. 1 Introduction Dependency-based Compositional Semantics (DCS) provides a formal yet intuitive way to model nat- ural language semantics. It was initially proposed in Liang et al. (2011) as a relational database query- ing protocol, and later used for logical inference in Tian et al. (2014a). Although the DCS inference framework provided decent support for both quanti- fiers all (universal quantifier) and no (negated exis- tential quantifier), attention is required for an RTE system to cope with generalized quantifiers (GQ), including “at most n”, “at least n”, “most”, etc., which can affect the direction or even the existence of an entailment relation, as demonstrated in Exam- ples 1 to 3. ⇤This work was conducted during an internship at the Na- tional Institute of Informatics, Japan. PExample 1. P ) H but H ; P, where At most 5 students like noodles. H At most 5 Japanese students like udon noodles. PExample 2. P ) H but H ; P, where At least 5 Japanese students like udon noodles. H At least 5 students like noodles. PExample 3. P ; H and H ; P, where Most Japanese students like udon noodles. H Most students like noodles. In this paper, we explore ways of encoding GQs in a recent framework of Dependency-based Com- positional Semantics (DCS) (Liang et al., 2013; Tian et al., 2014a), especially aiming to correctly handle linguistic knowledge like hyponymy when GQs are involved. We use selection operators, an extension mechanism described in Tian et al. (2014a), to im- plement a sub-type of GQs (Section 3.1). To deal with downward monotonicity of the predicate ar- gument, we also propose a simple extension called “relation” to the framework (Section 3.2). This ap- proach does not encode the exact semantics of ev- ery specific GQ, but instead captures some major properties that are both easily implementable with the current technology and useful in many cases. As in Tian et al. (2014a), we empirically tested the extended system on the “Generalized Quantifiers” section of the FraCaS corpus (The Fracas Consor- tium et al., 1996), and reduced 69% of the previous errors. A further error analysis reveals some limita- tions of the current approach, suggesting extensions towards more powerful logical systems. We hope this …

Content truncated to 500 words