Incorporating Contextual Information for Language-Independent, Dynamic Disambiguation Tasks

Incorporating Contextual Information for Language-Independent, Dynamic Disambiguation Tasks Tobias Staron, ¨Ozge Alac¸am, Wolfgang Menzel Department of Informatics University of Hamburg {staron, alacam, menzel}@informatik.uni-hamburg.de Abstract Humans resolve various kinds of linguistic ambiguities by exploiting available external evidence that has been acquired from modalities besides the linguistic one. This behavior can be observed for several languages, for English or German for example. In contrast, most natural language processing systems, parsers for example, rely on linguistic information only without taking further knowledge into account. While those systems are expected to correctly handle syntactically unambiguous cases, they cannot resolve syntactic ambiguities reliably. This paper hypothesizes that parsers would be able to find non-canonical interpretations of ambiguous sentences, if they exploited external, contextual information. The proposed multi-modal system, which combines data-driven and grammar-based approaches, confirmed this hypothesis in experiments on syntactically ambiguous sentences. This work focuses on the scarcely investigated relative clause attachment ambiguity instead of prepositional phrase attachment ambiguities, which are already well known in the literature. Experiments were conducted for English, German and Turkish and dynamic, i. e. highly dissimilar, contexts. Keywords: disambiguation, context, language-independence 1. Introduction In psycholinguistics, there is substantial empirical evidence suggesting that human language processing successfully integrates available information acquired from different modalities in order to resolve fully as well as temporally ambiguous linguistic input, e. g. on the syntactic level, and predict what will be revealed next in the unfolding sen- tence (Tanenhaus et al., 1995; Altmann and Kamide, 1999; Knoeferle, 2005). During spoken communication, disam- biguation and prediction processes allow for more accu- rate understanding. In contrast, natural language process- ing (NLP) systems are still not able to achieve that level of accuracy concerning challenging linguistic situations. For example, a parser that processes linguistic information is expected to successfully handle syntactically unambigu- ous sentences by applying knowledge derived from training data or linguistic rules. However, neither parsers nor hu- mans can resolve syntactic ambiguities without additional information. They may only have preferences. But, humans will use external information from other modalities for dis- ambiguation successfully if it becomes available. There- fore, we expect to resolve syntactic ambiguities via multi- modal disambiguation by exploiting external knowledge, i. e. contextual information, that is derived from another modality, e. g. from visual scenes of the described events. This work proposes two hypotheses. H1) A parser will resolve linguistic ambiguities reliably and reach correct interpretations if contextual infor- mation derived from additional modalities besides the linguistic one, i. e. visual scenes, are exploited. H2) In addition, this behavior is expected to be observed independent of the language. The contributions of this work are twofold. First, a lan- guage independent, data-driven parser has been modified to employ a grammar-based approach that will incorporate the contextual information even if it is previously unseen. Secondly, that system is used to validate the hypotheses for multiple languages: English, German and Turkish. One of the most frequently investigated cases of syntac- tic ambiguity are prepositional phrase (PP) attachment am- biguities, where different semantic and syntactic interpre- …

Content truncated to 500 words