Discourse-sensitive Automatic Identification of Generic Expressions

Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing, pages 1272–1281, Beijing, China, July 26-31, 2015. c⃝2015 Association for Computational Linguistics Discourse-sensitive Automatic Identification of Generic Expressions Annemarie Friedrich Manfred Pinkal Department of Computational Linguistics Saarland University, Saarbr¨ucken, Germany {afried,pinkal}@coli.uni-saarland.de Abstract This paper describes a novel sequence la- beling method for identifying generic ex- pressions, which refer to kinds or arbitrary members of a class, in discourse context. The automatic recognition of such expres- sions is important for any natural language processing task that requires text under- standing. Prior work has focused on iden- tifying generic noun phrases; we present a new corpus in which not only subjects but also clauses are annotated for generic- ity according to an annotation scheme mo- tivated by semantic theory. Our context- aware approach for automatically identi- fying generic expressions uses conditional random fields and outperforms previous work based on local decisions when evalu- ated on this corpus and on related data sets (ACE-2 and ACE-2005). 1 Introduction Distinguishing between statements about particu- lar individuals or situations and generic sentences is an important part of human language under- standing. Consider example (1): sentence (a) names characteristic attributes of a kind, which are inherent to every (typical) individual, and sentence (b) describes a specific individual. (1) (a) The modern domestic horse has a life expectancy of 25 to 30 years. (generic) (a) Old Billy lived to the age of 62. (non-generic) The above example illustrates that generic and non-generic sentences differ substantially in their semantic impact and entailment properties. It can be inferred from sentence (1a) that a typical horse has a life expectancy of 25 to 30 years, and if we know that Nelly is a horse, we can infer that its life expectancy is 25 to 30 years. Sentence (1b) has no such properties, it only allows inferences about the particular individual Old Billy. An automatic classifier that recognizes generic expressions would be extremely valuable for var- ious kinds of natural language processing sys- tems: for text understanding and question answer- ing systems, through the improvement of textual entailment methods, and for systems acquiring machine-readable knowledge from text. Machine- readable knowledge bases have different repre- sentations for statements corresponding to generic knowledge about kinds and knowledge about spe- cific individuals. The non-generic sentence (1b) roughly speaking provides ABox content for a machine-readable knowledge base, i.e., knowl- edge about particular instances, e.g, “A is an in- stance of B / has property X”. In contrast, the generic sentence (1a) feeds the TBox, i.e., knowl- edge of the form “All B are C / have property X”. Reiter and Frank (2010) provide a detailed discus- sion of the relevance of the distinction between classes and instances for automatic ontology con- struction. In this paper, we present a new corpus anno- tated in a linguistically motivated way for gener- icity, and a context-sensitive computational model for labeling sequences of clauses or noun phrases …

Content truncated to 500 words