Semantic Parsing with Relaxed Hybrid Trees

Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1308–1318, October 25-29, 2014, Doha, Qatar. c⃝2014 Association for Computational Linguistics Semantic Parsing with Relaxed Hybrid Trees Wei Lu Information Systems Technology and Design Singapore University of Technology and Design luwei@sutd.edu.sg Abstract We propose a novel model for parsing natural language sentences into their for- mal semantic representations. The model is able to perform integrated lexicon ac- quisition and semantic parsing, mapping each atomic element in a complete seman- tic representation to a contiguous word sequence in the input sentence in a re- cursive manner, where certain overlap- pings amongst such word sequences are allowed. It defines distributions over the novel relaxed hybrid tree structures which jointly represent both sentences and se- mantics. Such structures allow tractable dynamic programming algorithms to be developed for efficient learning and decod- ing. Trained under a discriminative set- ting, our model is able to incorporate a rich set of features where certain unbounded long-distance dependencies can be cap- tured in a principled manner. We demon- strate through experiments that by exploit- ing a large collection of simple features, our model is shown to be competitive to previous works and achieves state-of-the- art performance on standard benchmark data across four different languages. The system and code can be downloaded from http://statnlp.org/research/sp/. 1 Introduction Semantic parsing, the task of transforming natu- ral language sentences into formal representations of their underlying semantics, is one of the clas- sic goals for natural language processing and ar- tificial intelligence. This area of research recently has received a significant amount of attention. Var- ious models have been proposed over the past few years (Zettlemoyer and Collins, 2005; Kate and QUERY : answer(RIVER) RIVER : exclude(RIVER, RIVER) RIVER : traverse(STATE) STATE : stateid(STATENAME) STATENAME : (′tn′) RIVER : river(all) What rivers do not run through Tennessee ? Figure 1: An example tree-structured semantic representation (above) and its corresponding nat- ural language sentence. Mooney, 2006; Wong and Mooney, 2006; Lu et al., 2008; Jones et al., 2012). Following previous research efforts, we perform semantic parsing under a setting where the seman- tics for complete sentences are provided as train- ing data, but detailed word-level semantic infor- mation is not explicitly given during the training phase. As one example, consider the following natural language sentence paired with its corre- sponding semantic representation: What rivers do not run through Tennessee ? answer(exclude(river(all), traverse(stateid(′tn′)))) The training data consists of a set of sentences paired with semantic representations. Our goal is to learn from such pairs a model, which can be effectively used for parsing novel sentences into their semantic representations. Certain assumptions about the semantics are typically made. One common assumption is that the semantics can be represented as certain re- cursive structures such as trees, which consist of atomic semantic units as tree nodes. For exam- ple, the above semantics can be converted into an equivalent tree structure as illustrated in Figure 1. We will provide more details about such …

Content truncated to 500 words