Identification of Multi-word Expressions by Combining Multiple Linguistic Information Sources

Proceedings of the 2011 Conference on Empirical Methods in Natural Language Processing, pages 836–845, Edinburgh, Scotland, UK, July 27–31, 2011. c⃝2011 Association for Computational Linguistics Identification of Multi-word Expressions by Combining Multiple Linguistic Information Sources Yulia Tsvetkov Language Technologies Institute Carnegie Mellon University yulia.tsvetkov@gmail.com Shuly Wintner Department of Computer Science University of Haifa shuly@cs.haifa.ac.il Abstract We propose an architecture for expressing various linguistically-motivated features that help identify multi-word expressions in nat- ural language texts. The architecture com- bines various linguistically-motivated clas- sification features in a Bayesian Network. We introduce novel ways for computing many of these features, and manually de- fine linguistically-motivated interrelationships among them, which the Bayesian network models. Our methodology is almost en- tirely unsupervised and completely language- independent; it relies on few language re- sources and is thus suitable for a large num- ber of languages. Furthermore, unlike much recent work, our approach can identify ex- pressions of various types and syntactic con- structions. We demonstrate a significant im- provement in identification accuracy, com- pared with less sophisticated baselines. 1 Introduction Multi-word Expressions (MWEs) are lexical items that consist of multiple orthographic words (e.g., ad hoc, by and large, New York, kick the bucket). MWEs are numerous and constitute a significant portion of the lexicon of any natural language (Jack- endoff, 1997; Erman and Warren, 2000; Sag et al., 2002). They are a heterogeneous class of con- structions with diverse sets of characteristics, dis- tinguished by their idiosyncratic behavior. Mor- phologically, some MWEs allow some of their con- stituents to freely inflect while restricting (or pre- venting) the inflection of other constituents. In some cases MWEs may allow constituents to un- dergo non-standard morphological inflections that they would not undergo in isolation. Syntactically, some MWEs behave like words while other are phrases; some occur in one rigid pattern (and a fixed order), while others permit various syntactic trans- formations. Semantically, the compositionality of MWEs is gradual, ranging from fully compositional to idiomatic (Bannard et al., 2003). Because of their prevalence and irregularity, MWEs must be stored in lexicons of natural lan- guage processing applications. Correct handling of MWEs has been proven beneficial for various ap- plications, including information retrieval, building ontologies, text alignment, and machine translation. We propose a novel architecture for identifying MWEs of various types and syntactic categories in monolingual corpora. Unlike much existing work, which focuses on a particular syntactic construction, our approach addresses MWEs of all types by focus- ing on the general idiosyncratic properties of MWEs rather than on specific properties of each sub-class thereof. While we only evaluate our methodol- ogy on bi-grams, it can in principle be extended to longer MWEs. The architecture uses Bayesian Networks (BN) to express multiple interdependent linguistically-motivated features. First, we automatically generate a small (training) set of MWE and non-MWE bi-grams (positive and negative instances, respectively). We then define a set of linguistically-motivated features that embody observed characteristics of MWEs. We augment these by features that reflect collocation measures. Finally, we define dependencies among these …

Content truncated to 500 words