Semantic Kernels for Semantic Parsing

Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 436–442, October 25-29, 2014, Doha, Qatar. c⃝2014 Association for Computational Linguistics Semantic Kernels for Semantic Parsing Iman Saleh Faculty of Computers and Information Cairo University iman.saleh@fci-cu.edu.eg Alessandro Moschitti, Preslav Nakov, Llu´ıs M`arquez, Shafiq Joty ALT Research Group Qatar Computing Research Institute {amoschitti,pnakov,lmarquez,sjoty}@qf.org.qa Abstract We present an empirical study on the use of semantic information for Concept Seg- mentation and Labeling (CSL), which is an important step for semantic parsing. We represent the alternative analyses out- put by a state-of-the-art CSL parser with tree structures, which we rerank with a classifier trained on two types of seman- tic tree kernels: one processing structures built with words, concepts and Brown clusters, and another one using semantic similarity among the words composing the structure. The results on a corpus from the restaurant domain show that our semantic kernels exploiting similarity measures out- perform state-of-the-art rerankers. 1 Introduction Spoken Language Understanding aims to inter- pret user utterances and to convert them to logical forms or, equivalently, to database queries, which can then be used to satisfy the user’s information needs. This process is known as Concept Segmen- tation and Labeling (CSL), also called semantic parsing in the speech community: it maps utter- ances into meaning representations based on se- mantic constituents. The latter are basically word sequences, often referred to as concepts, attributes or semantic tags. CSL makes it easy to convert spoken questions such as “cheap lebanese restau- rants in doha with take out” into database queries. First, a language-specific semantic parser tok- enizes, 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 mislabeled segments: [Price low] [Cuisine lebanese] [City doha] [Amenity carry out] Finally, a database query is formed from the list of labels and values, and is then executed against the database, e.g., MongoDB; a backoff mecha- nism may be used if the query has not succeeded. {$and [{cuisine:"lebanese"},{city:"doha"}, {price:"low"},{amenity:"carry out"}]} The state-of-the-art of CSL is represented by conditional models for sequence labeling such as Conditional Random Fields (CRFs) (Lafferty et al., 2001) trained with simple morphological and lexical features. The basic CRF model was im- proved by means of reranking (Moschitti et al., 2006; Dinarelli et al., 2012) using structural ker- nels (Moschitti, 2006). Although these meth- ods exploited sentence structure, they did not use syntax at all. More recently, we applied shal- low syntactic structures and discourse parsing with slightly better results (Saleh et al., 2014). How- ever, the most obvious models for semantic pars- ing, i.e., rerankers based on semantic structural kernels (Bloehdorn and Moschitti, 2007b), had not been applied to semantic structures yet. In this paper, we study the impact of semantic information conveyed by Brown Clusters (BCs) (Brown et al., 1992) and semantic similarity, while also combining them with innovative features. We use reranking, similarly to (Saleh et …

Content truncated to 500 words