Learning to Make Inferences in a Semantic Parsing Task

Learning to Make Inferences in a Semantic Parsing Task Kyle Richardson and Jonas Kuhn Institut f¨ur Maschinelle Sprachverarbeitung University of Stuttgart {kyle,jonas.kuhn}@ims.uni-stuttgart.de Abstract We introduce a new approach to training a semantic parser that uses textual entail- ment judgements as supervision. These judgements are based on high-level in- ferences about whether the meaning of one sentence follows from another. When applied to an existing semantic parsing task, they prove to be a useful tool for revealing semantic distinctions and back- ground knowledge not captured in the tar- get representations. This information is used to improve the quality of the seman- tic representations being learned and to acquire generic knowledge for reasoning. Experiments are done on the benchmark Sportscaster corpus (Chen and Mooney, 2008), and a novel RTE-inspired infer- ence dataset is introduced. On this new dataset our method strongly outperforms several strong baselines. Separately, we obtain state-of-the-art results on the orig- inal Sportscaster semantic parsing task. 1 Introduction Semantic Parsing is the task of automatically translating natural language text to formal mean- ing representations (e.g., statements in a formal logic). Recent work has centered around learning such translations using parallel data, or raw col- lections of text-meaning pairs, often by employ- ing methods from statistical machine translation (Wong and Mooney, 2006; Jones et al., 2012; An- dreas et al., 2013) and parsing (Zettlemoyer and Collins, 2009; Kwiatkowski et al., 2010). Ear- lier attempts focused on learning to map natural language questions to simple database queries for database retrieval using collections of target ques- tions and formal queries. A more recent focus has been on learning representations using weaker forms of supervision that require minimal amounts of manual annotation effort (Clarke et al., 2010; Liang et al., 2011; Krishnamurthy and Mitchell, 2012; Artzi and Zettlemoyer, 2013; Berant et al., 2013; Kushman et al., 2014). For example, Liang et al. (2011) train a seman- tic parser in a question-answering domain using the denotation (or answer) of each question as the sole supervision. Particularly impressive is their system’s ability to learn complex linguistic struc- ture not handled by earlier methods that use more direct supervision. Similarly, Artzi and Zettle- moyer (2013) train a parser that generates higher- order logical representations in a navigation do- main using low-level navigation cues. What is missing in such approaches, however, is an ex- plicit account of entailment (e.g., learning entail- ment rules from such corpora), which has long been considered one of the basic aims of se- mantics (Montague, 1970). An adequate seman- tic parser that captures the core aspects of natu- ral language meaning should support inferences about sentence-level entailments (i.e., determining whether the meaning of one sentence follows from another). In many cases, the target representations being learned remain inexpressive, making it diffi- cult to learn the types of semantic generalizations and world-knowledge needed for modeling entail- ment (see discussion in Schubert (2015)). Attempts to integrate more general knowledge into semantic parsing pipelines have often in- volved additional hand-engineering or external lexical resources (Wang et …

Content truncated to 500 words