A Study of using Syntactic and Semantic Structures for Concept Segmentation and Labeling

Proceedings of COLING 2014, the 25th International Conference on Computational Linguistics: Technical Papers, pages 193–202, Dublin, Ireland, August 23-29 2014. A Study of using Syntactic and Semantic Structures for Concept Segmentation and Labeling Iman Saleh∗, Shafiq Joty, Llu´ıs M`arquez, Alessandro Moschitti, Preslav Nakov ALT Research Group Qatar Computing Research Institute {sjoty,lmarquez,amoschitti,pnakov} @qf.org.qa Scott Cyphers, Jim Glass MIT CSAIL Cambridge, Massachusetts 02139 USA {cyphers,glass}@mit.edu Abstract This paper presents an empirical study on using syntactic and semantic information for Concept Segmentation and Labeling (CSL), a well-known component in spoken language understand- ing. Our approach is based on reranking N-best outputs from a state-of-the-art CSL parser. We perform extensive experimentation by comparing different tree-based kernels with a variety of representations of the available linguistic information, including semantic concepts, words, POS tags, shallow and full syntax, and discourse trees. The results show that the structured representa- tion with the semantic concepts yields significant improvement over the base CSL parser, much larger compared to learning with an explicit feature vector representation. We also show that shallow syntax helps improve the results and that discourse relations can be partially beneficial. 1 Introduction Spoken Language Understanding aims to interpret user utterances and to convert them to logical forms, or, equivalently, database queries, which can then be used to satisfy the user’s information needs. This process is known as Concept Segmentation and Labeling (CSL): it maps utterances into meaning repre- sentations based on semantic constituents. The latter are basically sequences of semantic entities, often referred to as concepts, attributes or semantic tags. Traditionally, grammar-based methods have been used for CSL, but more recently machine learning approaches to semantic structure computation have been shown to yield higher accuracy. However, most previous work did not exploit syntactic/semantic structures of the utterances, and the state-of-the-art is represented by conditional models for sequence la- beling, such as Conditional Random Fields (Lafferty et al., 2001) trained with simple morphological and lexical features. In our study, we measure the impact of syntactic and discourse structures by also com- bining them with innovative features. In the following subsections, we present the application context for our CSL task and then we outline the challenges and the findings of our research. 1.1 Semantic parsing for the “restaurant” domain We experiment with the dataset of McGraw et al. (2012), containing spoken and typed questions about restaurants, which are to be answered using a database of free text such as reviews, categorical data such as names and locations, and semi-categorical data such as user-reported cuisines and amenities. Semantic parsing, in the form of sequential segmentation and labeling, makes it easy to convert spoken and typed questions such as “cheap lebanese restaurants in doha with take out” into database queries. First, a language-specific semantic parser tokenizes, segments and labels the question: [Price cheap] [Cuisine lebanese] [Other restaurants in] [City doha] [Other with] [Amenity take out] Then, label-specific normalizers are applied to the segments, with the option to possibly relabel mis- labeled segments; at this point, discourse history may be incorporated …

Content truncated to 500 words