A Methodology for Automatic Identification of Nocuous Ambiguity

Proceedings of the 23rd International Conference on Computational Linguistics (Coling 2010), pages 1218–1226, Beijing, August 2010 A Methodology for Automatic Identification of Nocuous Ambiguity Hui Yang1 Anne de Roeck1 Alistair Willis1 Bashar Nuseibeh1, 2 1Department of Computing, The Open University 2Lero, University of Limerick {h.yang, a.deroeck, a.g.willis, b.nuseibeh}@open.ac.uk Abstract Nocuous ambiguity occurs when a lin- guistic expression is interpreted differ- ently by different readers in a given con- text. We present an approach to auto- matically identify nocuous ambiguity that is likely to lead to misunderstand- ings among readers. Our model is built on a machine learning architecture. It learns from a set of heuristics each of which predicts a factor that may lead a reader to favor a particular interpretation. An ambiguity threshold indicates the ex- tent to which ambiguity can be tolerated in the application domain. Collections of human judgments are used to train heu- ristics and set ambiguity thresholds, and for evaluation. We report results from applying the methodology to coordina- tion and anaphora ambiguity. Results show that the method can identify nocu- ous ambiguity in text, and may be wid- ened to cover further types of ambiguity. We discuss approaches to evaluation. 1 Introduction Traditional accounts of ambiguity have generally assumed that each use of a linguistic expression has a unique intended interpretation in context, and attempted to develop a model to determine it (Nakov and Hearst, 2005; Brill and Resnik, 1994). However, disambiguation is not always appropriate or even desirable (Poesio and Art- stein, 2008). Ambiguous text may be interpreted differently by different readers, with no consen- sus about which reading is the intended one. At- tempting to assign a preferred interpretation may therefore be inappropriate. Misunderstandings among readers do occur and may have undesir- able consequences. In requirements engineering processes, for example, this results in costly im- plementation errors (Boyd et al., 2005). Nonetheless, most text does not lead to sig- nificant misinterpretation. Our research aims to establish a model that estimates how likely an ambiguity is to lead to misunderstandings. Our previous work on nocuous ambiguity (Chantree et al., 2006; Willis et al., 2008) cast ambiguity not as a property of a text, but as a property of text in relation to a set of stakeholders. We drew on human judgments - interpretations held by a group of readers of a text – to establish criteria for judging the presence of nocuous ambiguity. An ambiguity is innocuous if it is read in the same way by different people, and nocuous oth- erwise. The model was tested on co-ordination ambiguity only. In this paper, we implement, refine and extend the model. We investigate two typical ambiguity types arising from coordination and anaphora. We extend the previous work (Willis et al., 2008) with additional heuristics, and refine the concept of ambiguity threshold. We experiment with alternative machine learning algorithms to find optimal ways of combining the output of the heuristics. Yang et al. (2010a) describes a com- plete implementation in a prototype tool running on …

Content truncated to 500 words