Improving the Quality of Text Understanding by Delaying Ambiguity Resolution

Proceedings of the 23rd International Conference on Computational Linguistics (Coling 2010), pages 581–589, Beijing, August 2010 Improving the Quality of Text Understanding by Delaying Ambiguity Resolution Doo Soon Kim Dept. of Computer Science University of Texas onue5@cs.utexas.edu Ken Barker Dept. of Computer Science University of Texas kbarker@cs.utexas.edu Bruce Porter Dept. of Computer Science University of Texas porter@cs.utexas.edu Abstract Text Understanding systems often commit to a single best interpretation of a sen- tence before analyzing subsequent text. This interpretation is chosen by resolv- ing ambiguous alternatives to the one with the highest confidence, given the context available at the time of commitment. Sub- sequent text, however, may contain infor- mation that changes the confidence of al- ternatives. This may especially be the case with multiple redundant texts on the same topic. Ideally, systems would de- lay choosing among ambiguous alterna- tives until more text has been read. One solution is to maintain multiple can- didate interpretations of each sentence un- til the system acquires disambiguating ev- idence. Unfortunately, the number of al- ternatives explodes quickly. In this pa- per, we propose a packed graphical (PG) representation that can efficiently repre- sent a large number of alternative interpre- tations along with dependencies among them. We also present an algorithm for combining multiple PG representations to help resolve ambiguity and prune alterna- tives when the time comes to commit to a single interpretation. Our controlled experiments show that by delaying ambiguity resolution until multi- ple texts have been read, our prototype’s accuracy is higher than when committing to interpretations sentence-by-sentence. 1 Introduction A typical text understanding system confronts am- biguity while parsing, mapping words to concepts and formal relations, resolving co-references, and integrating knowledge derived from separate sen- tences or texts. The system discards many candi- date interpretations to avoid combinatorial explo- sion. Commonly, after reading each sentence, a system will commit to its top ranked interpreta- tion of the sentence before reading the next. If a text understanding system could postpone committing to an interpretation without being swamped by a combinatorial explosion of alterna- tives, its accuracy would almost surely improve. This intuition follows from the observation that text is redundant in at least two ways. First, within a single coherent text (about the same entities and events), each sentence informs the interpre- tation of its neighbors. Second, within a corpus of texts on the same topic, the same information is expressed in different surface forms, ambiguous in different ways. Related fields, such as Infor- mation Extraction, exploit textual redundancy to good effect, and perhaps text understanding can as well. One approach is for the text understanding sys- tem to maintain multiple complete candidate in- terpretations. After reading each sentence, for ex- ample, the system would retain a beam of the n- best interpretations of the sentence. While this approach avoids a combinatorial explosion (for reasonable values of n), several problems remain. First, because the beam width is limited, the sys- tem may still discard correct interpretations before benefiting from the extra …

Content truncated to 500 words