The Role of Context in Neural Morphological Disambiguation

Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers, pages 181–191, Osaka, Japan, December 11-17 2016. The Role of Context in Neural Morphological Disambiguation Qinlan Shen, Daniel Clothiaux, Emily Tagtow, Patrick Littell, Chris Dyer Carnegie Mellon University {qinlans, dclothia, etagtow, plittell, cdyer}@cs.cmu.edu Abstract Languages with rich morphology often introduce sparsity in language processing tasks. While morphological analyzers can reduce this sparsity by providing morpheme-level analyses for words, they will often introduce ambiguity by returning multiple analyses for the same surface form. The problem of disambiguating between these morphological parses is further compli- cated by the fact that a correct parse for a word is not only dependent on the surface form but also on other words in its context. In this paper, we present a language-agnostic approach to morphological disambiguation. We address the problem of using context in morphological disambiguation by presenting several LSTM-based neural architectures that encode long-range surface-level and analysis-level contextual dependencies. We applied our approach to Turkish, Russian, and Arabic to compare effectiveness across languages, matching state-of-the-art in two of the three languages. Our results also demonstrate that while context plays a role in learning how to disambiguate, the type and amount of context needed varies between languages based on their morphological and syntactic properties. 1 Introduction Morphologically rich languages introduce sparsity in language processing tasks, as different surface variants over the same root are often taken as independent entities. Using a morphological analyzer can decompose inflected words into known tags that encode syntactic and semantic information about the word. However, finding the correct morphological parse is a non-trivial task. Functionally different morphemes may have similar forms, and long strings of potentially ambiguous morphemes compound the problem of ambiguity. As a result, analyzers for morphologically complex languages often return several parses for the same surface word. Table 1, for example, shows the resulting candidate parses for the surface form “alın” returned by Oflazer’s (1994) morphological analyzer for Turkish. alın+Noun+A3sg+Pnon+Nom (forehead) al+AdjˆDB+Noun+Zero+A3sg+P2sg+Nom (your red) al+AdjˆDB+Noun+Zero+A3sg+Pnon+Gen (of red) al+Verb+Pos+Imp+A2pl ((you) take) al+VerbˆDB+Verb+Pass+Pos+Imp+A2sg ((you) be taken) alın+Verb+Pos+Imp+A2sg ((you) be offended) Table 1: Possible morphological parses for surface form “alın” The surrounding context of the word being disambiguated plays a major part in determining the role of a word in a sentence and thus its correct morphological parse. For example, the surface form “evi” in Turkish can be interpreted as either accusative or third-person singular possessive, as shown in Table 2 (Yildiz et al., 2016); this determination cannot reliably be made without further context. Moreover, the disambiguation of a word can in turn disambiguate other words; if we have determined that “evi” is This work is licensed under a Creative Commons Attribution 4.0 International Licence. Licence details: http://creativecommons.org/licenses/by/4.0/ 181 accusative, we can infer that a transitive verb must follow and that other ambiguous forms in a sentence are not accusative. Sentence and translation Analysis of evi Evi bulabildiniz mi? – Did ev+Noun+3sg you find the house? +Pnon+Acc Evi gerc¸ekten g¨uzelmis¸. – ev+Noun+3sg His/Her house is really +P3sg+Nom …

Content truncated to 500 words