Does Free Word Order Hurt? Assessing the Practical Lexical Function Model for Croatian

Proceedings of the 6th Joint Conference on Lexical and Computational Semantics (*SEM 2017), pages 115–120, Vancouver, Canada, August 3-4, 2017. c⃝2017 Association for Computational Linguistics Does Free Word Order Hurt? Assessing the Practical Lexical Function Model for Croatian Zoran Medi´c∗ Jan Šnajder∗ Sebastian Padó† ∗ Faculty of Electrical Engineering and Computing, University of Zagreb {jan.snajder, zoran.medic}@fer.hr † Institut für Maschinelle Sprachverarbeitung, Stuttgart University pado@ims.uni-stuttgart.de Abstract The Practical Lexical Function (PLF) model is a model of computational distri- butional semantics that attempts to strike a balance between expressivity and learn- ability in predicting phrase meaning and shows competitive results. We investigate how well the PLF carries over to free word order languages, given that it builds on ob- servations of predicate-argument combina- tions that are harder to recover in free word order languages. We evaluate variants of the PLF for Croatian, using a new lexical substitution dataset. We find that the PLF works about as well for Croatian as for En- glish, but demonstrate that its strength lies in modeling verbs, and that the free word order affects the less robust PLF variant. 1 Introduction Compositional distributional semantic models (CDSMs) represent phrase meaning in a vector space by composing the meanings of individual words. Many CDSMs were proposed, ranging from basic ones that use element-wise operations on word vectors to compute phrase vectors (Mitchell and Lapata, 2008), to more complex models that represent predicate arguments as higher-order ten- sors (Baroni and Zamparelli, 2010; Guevara, 2010). The latter models assume that predicates in a phrase act as functions that act on other phrase compo- nents to yield the final representation of the phrase. For example, an adjective acts as a function on the noun in an adjective-noun phrase, while a transitive verb acts as a binary function on its subject and object. However, since the number of parameters in a tensor grows exponentially with the number of arguments of the function that it models, learning full tensors for predicates with many arguments is tedious to impractical (Grefenstette et al., 2012). The Practical Lexical Function model (PLF, Pa- perno et al. (2014)) strikes a middle ground by breaking down all tensors with ranks higher than two into multiple matrices, each representing the predicate’s composition with a single argument (cf. Section 2 for details). In the experiments of Paperno et al. (2014), PLF has been shown to work better than some other CDSMs in modeling seman- tic similarity. Particularly good results were ob- tained on ANVAN (adjective-noun-verb-adjective- noun) phrases, where PLF outperformed both sim- ple CDSMs (due to its higher expressiveness) as well as the higher-order Lexical Function model (Baroni and Zamparelli, 2010). Although the PLF shows promising results, ex- isting work still leaves open two questions. First, it is not obvious that these results carry over to languages with free word order, such as Slavic lan- guages, where predicates and arguments are often separated. For example, in the English sentence ‘I like my dog’, the predicate is adjacent to both the subject and …

Content truncated to 500 words