Language-Independent Discriminative Parsing of Temporal Expressions

Proceedings of the 51st Annual Meeting of the Association for Computational Linguistics, pages 83–92, Sofia, Bulgaria, August 4-9 2013. c⃝2013 Association for Computational Linguistics Language-Independent Discriminative Parsing of Temporal Expressions Gabor Angeli Stanford University Stanford, CA 94305 angeli@stanford.edu Jakob Uszkoreit Google, Inc. 1600 Amphitheatre Parkway Mountain View, CA 94303 uszkoreit@google.com Abstract Temporal resolution systems are tradition- ally tuned to a particular language, re- quiring significant human effort to trans- late them to new languages. We present a language independent semantic parser for learning the interpretation of tempo- ral phrases given only a corpus of utter- ances and the times they reference. We make use of a latent parse that encodes a language-flexible representation of time, and extract rich features over both the parse and associated temporal semantics. The parameters of the model are learned using a weakly supervised bootstrapping approach, without the need for manually tuned parameters or any other language expertise. We achieve state-of-the-art ac- curacy on all languages in the TempEval- 2 temporal normalization task, reporting a 4% improvement in both English and Spanish accuracy, and to our knowledge the first results for four other languages. 1 Introduction Temporal resolution is the task of mapping from a textual phrase describing a potentially complex time, date, or duration to a normalized (grounded) temporal representation. For example, possibly complex phrases such as the week before last1 are often more useful in their grounded form – e.g., August 4 - August 11. Many approaches to this problem make use of rule-based methods, combining regular- expression matching and hand-written interpreta- tion functions. In contrast, we would like to learn the interpretation of a temporal expression proba- bilistically. This allows propagation of uncertainty to higher-level components, and the potential to 1Spoken on, for instance, August 20. dynamically back off to a rule-based system in the case of low confidence parses. In addition, we would like to use a representation of time which is broadly applicable to multiple languages, without the need for language-specific rules or manually tuned parameters. Our system requires annotated data consist- ing only of an input phrase and an associ- ated grounded time, relative to some reference time; the language-flexible parse is entirely latent. Training data of this weakly-supervised form is generally easier to collect than the alternative of manually creating and tuning potentially complex interpretation rules. A large number of languages conceptualize time as lying on a one dimensional line. Although the surface forms of temporal expressions differ, the basic operations many languages use can be mapped to operations on this time line (see Sec- tion 3). Furthermore, many common languages share temporal units (hours, weekdays, etc.). By structuring a latent parse to reflect these seman- tics, we can define a single model which performs well on multiple languages. A discriminative parsing model allows us to de- fine sparse features over not only lexical cues but also the temporal value of our prediction. For ex- ample, it allows us to learn that we are much more likely to express …

Content truncated to 500 words