The Effects of Syntactic Features in Automatic Prediction of Morphology

Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing, pages 333–344, Seattle, Washington, USA, 18-21 October 2013. c⃝2013 Association for Computational Linguistics The Effects of Syntactic Features in Automatic Prediction of Morphology Wolfgang Seeker and Jonas Kuhn Institute for Natural Language Processing University of Stuttgart {seeker,jonas}@ims.uni-stuttgart.de Abstract Morphology and syntax interact considerably in many languages and language processing should pay attention to these interdependen- cies. We analyze the effect of syntactic fea- tures when used in automatic morphology pre- diction on four typologically different lan- guages. We show that predicting morphology for languages with highly ambiguous word forms profits from taking the syntactic context of words into account and results in state-of- the-art models. 1 Introduction In this paper, we investigate the interplay between syntax and morphology with respect to the task of assigning morphological descriptions (or tags) to each token of a sentence. Specifically, we examine the effect of syntactic information when it is inte- grated into the feature model of a morphological tag- ger. We test the effect of syntactic features on four languages – Czech, German, Hungarian, and Span- ish – and find that syntactic features improve our tag- ger considerably for Czech and German, but not for Hungarian and Spanish. Our analysis of construc- tions that show morpho-syntactic agreement sug- gests that syntactic features are important if the lan- guage shows frequent word form syncretisms1 that can be disambiguated by the syntactic context. The meaning of a sentence is structurally encoded 1Syncretism describes the situation where a word form is ambiguous between several different morphological descrip- tions within its inflection paradigm. by morphological and syntactic means.2 Different languages, however, use them to a different extent. Languages like English encode grammatical infor- mation (like the subject vs object status of an argu- ment) via word order, whereas languages like Czech or Hungarian use different word forms. Automatic analysis of languages with rich morphology needs to pay attention to the interaction between morphol- ogy and syntax in order to arrive at suitable com- putational models. Linguistic theory (e. g., Bresnan (2001), Melˇcuk (2009)) suggests many interactions between morphology and syntax. For example, lan- guages with a case system use different forms of the same word to mark different syntactic (or seman- tic) relations (Blake, 2001). In many languages, two words that participate in a syntactic relation show covariance in some or all of their morphological fea- tures (so-called agreement, Corbett (2006)).3 Automatic annotation of morphology assigns morphological descriptions (e. g., nominative- singular-masculine) to word forms. It is usually modeled as a sequence model, often in combination with part-of-speech tagging and lemmatization (Collins, 2002; Hajiˇc, 2004; Smith et al., 2005; Chrupała et al., 2008, and others). Sequence models achieve high accuracy and coverage but since they only use linear context they only approximate some of the underlying hierarchical relationships. As an example for these hierarchical relationships, 2And also by prosodic means, which we will not discuss since text-based tools rarely have access to this information. 3For …

Content truncated to 500 words