Identifying 1950s American Jazz Musicians: Fine-Grained IsA Extraction via Modifier Composition

Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics, pages 2099–2109 Vancouver, Canada, July 30 - August 4, 2017. c⃝2017 Association for Computational Linguistics https://doi.org/10.18653/v1/P17-1192 Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics, pages 2099–2109 Vancouver, Canada, July 30 - August 4, 2017. c⃝2017 Association for Computational Linguistics https://doi.org/10.18653/v1/P17-1192 Identifying 1950s American Jazz Musicians: Fine-Grained IsA Extraction via Modifier Composition Ellie Pavlick∗ University of Pennsylvania 3330 Walnut Street Philadelphia, Pennsylvania 19104 epavlick@seas.upenn.edu Marius Pa¸sca Google Inc. 1600 Amphitheatre Parkway Mountain View, California 94043 mars@google.com Abstract We present a method for populating fine-grained classes (e.g., “1950s Amer- ican jazz musicians”) with instances (e.g., Charles Mingus). While state- of-the-art methods tend to treat class labels as single lexical units, the pro- posed method considers each of the in- dividual modifiers in the class label rel- ative to the head. An evaluation on the task of reconstructing Wikipedia cate- gory pages demonstrates a >10 point increase in AUC, over a strong baseline relying on widely-used Hearst patterns. 1 Introduction The majority of approaches (Snow et al., 2006; Shwartz et al., 2016) for extracting IsA rela- tions from text rely on lexical patterns as the primary signal of whether an instance belongs to a class. For example, observing a pattern like “X such as Y” is a strong indication that Y (e.g., “Charles Mingus”) is an instance of class X (e.g., “musician”) (Hearst, 1992). Methods based on these “Hearst patterns” assume that class labels can be treated as atomic lexicalized units. This assumption has several significant weakness. First, in order to recognize an instance of a class, these pattern- based methods require that the entire class la- bel be observed verbatim in text. The require- ment is reasonable for class labels containing a single word, but in practice, there are many possible fine-grained classes: not only “mu- sicians” but also “1950s American jazz mu- sicians”. The probability that a given label will appear in its entirety within one of the expected patterns is very low, even in large ∗Contributed during an internship at Google. 1950s American jazz musicians . . . seminal musicians such as Charles Mingus and George Russell. . . . . . A virtuoso bassist and composer, Mingus ir- revocably changed the face of jazz. . . . . . Mingus truly was a product of America in all its historic complexities. . . . . . Mingus dominated the scene back in the 1950s and 1960s. . . Figure 1: We extract instances of fine-grained classes by considering each of the modifiers in the class label individually. This allows us to extract instances even when the full class label never appears in text. amounts of text. Second, when class labels are treated as though they cannot be decom- posed, every class label must be modeled inde- pendently, even those containing overlapping words (“American jazz musician”, “French jazz musician”). As a result, the number of meaning representations to be learned is ex- ponential in …

Content truncated to 500 words