Context-dependent Semantic Parsing for Time Expressions

Proceedings of the 52nd Annual Meeting of the Association for Computational Linguistics, pages 1437–1447, Baltimore, Maryland, USA, June 23-25 2014. c⃝2014 Association for Computational Linguistics Context-dependent Semantic Parsing for Time Expressions Kenton Lee†, Yoav Artzi†, Jesse Dodge‡∗, and Luke Zettlemoyer† † Computer Science & Engineering, University of Washington, Seattle, WA {kentonl, yoav, lsz}@cs.washington.edu ‡ Language Technologies Institute, Carnegie Mellon University, Pittsburgh, PA jessed@cs.cmu.edu Abstract We present an approach for learning context-dependent semantic parsers to identify and interpret time expressions. We use a Combinatory Categorial Gram- mar to construct compositional meaning representations, while considering contex- tual cues, such as the document creation time and the tense of the governing verb, to compute the final time values. Exper- iments on benchmark datasets show that our approach outperforms previous state- of-the-art systems, with error reductions of 13% to 21% in end-to-end performance. 1 Introduction Time expressions present a number of challenges for language understanding systems. They have rich, compositional structure (e.g., “2nd Friday of July”), can be easily confused with non-temporal phrases (e.g., the word “May” can be a month name or a verb), and can vary in meaning in dif- ferent linguistic contexts (e.g., the word “Friday” refers to different dates in the sentences “We met on Friday” and “We will meet on Friday”). Recov- ering the meaning of time expressions is therefore challenging, but provides opportunities to study context-dependent language use. In this paper, we present the first context-dependent semantic pars- ing approach for learning to identify and interpret time expressions, addressing all three challenges. Existing state-of-the-art methods use hand- engineered rules for reasoning about time expres- sions (Str¨otgen and Gertz, 2013). This includes both detection, identifying a phrase as a time ex- pression, and resolution, mapping such a phrase into a standardized time value. While rule-based approaches provide a natural way to express ex- pert knowledge, it is relatively difficult to en- ∗Work conducted at the University of Washington. code preferences between similar competing hy- potheses and provide prediction confidence. Re- cently, methods for learning probabilistic seman- tic parsers have been shown to address such limi- tations (Angeli et al., 2012; Angeli and Uszkoreit, 2013). However, these approaches do not account for any surrounding linguistic context and were mainly evaluated with gold standard mentions. We propose to use a context-dependent se- mantic parser for both detection and resolution of time expressions. For both tasks, we make use of a hand-engineered Combinatory Catego- rial Grammar (CCG) to construct a set of mean- ing representations that identify the time being described. For example, this grammar maps the phrase “2nd Friday of July” to the meaning repre- sentation intersect(nth(2, friday), july), which encodes the set of all such days. Detection is then performed with a binary classifier to prune the set of text spans that can be parsed with the gram- mar (e.g., to tell that “born in 2000” has a time expression but “a 2000 piece puzzle” does not). For resolution, we consider mentions sequentially and use a log-linear model to select …

Content truncated to 500 words