Learning to Resolve Bridging References

LEARNING TO RESOLVE BRIDGING REFERENCES Massimo Poesio,♣ Rahul Mehta,♣ Axel Maroudas,♣ and Janet Hitzeman♠ ♣Dept. of Comp. Science, University of Essex, UK poesio at essex dot ac dot uk ♠MITRE Corporation, USA hitz at mitre dot org Abstract We use machine learning techniques to find the best combination of local focus and lexical distance features for identifying the anchor of mereological bridging references. We find that using first men- tion, utterance distance, and lexical distance com- puted using either Google or WordNet results in an accuracy significantly higher than obtained in pre- vious experiments. 1 Introduction BRIDGING REFERENCES (BR) (Clark, 1977)– anaphoric expressions that cannot be resolved purely on the basis of string matching and thus re- quire the reader to ’bridge’ the gap using common- sense inferences–are arguably the most interesting and, at the same time, the most challenging prob- lem in anaphora resolution. Work such as (Poesio et al., 1998; Poesio et al., 2002; Poesio, 2003) pro- vided an experimental confirmation of the hypoth- esis first put forward by Sidner (1979) that BRIDG- ING DESCRIPTIONS (BD)1 are more similar to pro- nouns than to other types of definite descriptions, in that they are sensitive to the local rather than the global focus (Grosz and Sidner, 1986). This previ- uous work also suggested that simply choosing the entity whose description is lexically closest to that of the bridging description among those in the cur- rent focus space gives poor results; in fact, better re- sults are obtained by always choosing as ANCHOR of the bridging reference2 the first-mentioned entity of the previous sentence (Poesio, 2003). But nei- ther source of information in isolation resulted in an accuracy over 40%. In short, this earlier work sug- gested that a combination of salience and lexical / 1We will use the term bridging descriptions to indicate bridging references realized by definite descriptions, equated here with noun phrases with determiner the, like the top. 2Following (Poesio and Vieira, 1998), we use the term ‘an- chor’ as as a generalization of the term ANTECEDENT, to indi- cate the discourse entity which an anaphoric expression either realizes, or is related to by an associative relation; reserving ‘antecedent’ for the cases of identity. commonsense information is needed to choose the most likely anchor; the problem remained of how to combine this information. In the work described in this paper, we used ma- chine learning techniques to find the best combina- tion of local focus features and lexical distance fea- tures, focusing on MEREOLOGICAL bridging refer- ences:3 references referring to parts of an object al- ready introduced (the cabinet), such as the panels or the top (underlined) in the following example from the GNOME corpus (Poesio et al., 2004). (1) The combination of rare and expensive ma- terials used on [this cabinet]i indicates that it was a particularly expensive commission. The four Japanese lacquer panels date from the mid- to late 1600s and were created with a technique known as kijimaki-e. For this type of lacquer, artisans …

Content truncated to 500 words