Pronominal anaphora resolution in KANTOO English-to-Spanish machine translation system

Pronominal Anaphora Resolution in KANTOO English-to-Spanish Machine Translation System Teruko Mitamura, Eric Nyberg, Enrique Torrejon, David Svoboda, Kathryn Baker Language Technologies Institute Carnegie Mellon University Pittsburgh, PA 15213 teruko@cs.cmu.edu Abstract We describe the automatic resolution of pronominal anaphora using KANT Controlled English (KCE) and the KANTOO English-to-Spanish MT system. Our algorithm is based on a robust, syntax-based approach that applies a set of restrictions and preferences to select the correct antecedent. We report a success rate of 89.6% on a training corpus with 289 anaphors, and 87.5% on held-out data containing 145 anaphors. Resolution of anaphors is important in translation, due to gender mismatches among languages; our approach translates anaphors to Spanish with 97.2% accuracy. Keywords Pronominal Anaphora Resolution, English to Spanish Machine Translation Introduction It has been claimed that encoding all the linguistic and domain knowledge required for successful pronominal anaphora resolution is a difficult and time-consuming task. Recent research in anaphora resolution seeks a simple, rapid, yet reliable approach which does not require extensive syntactic and semantic knowledge (Mitkov, 1998; Nasukawa 1994). In this work, we explore the extension of KANT Controlled English to include pronominal anaphors, and present an algorithm that effectively resolves anaphors while preserving the accuracy of the translated text. In contrast with statistical methods, our approach does not require a large bilingual aligned corpus for training. Our resolution algorithm is syntax-based, but we also follow the trend toward a simple, rapid approach. Our approach draws from linguistic approaches in earlier work (Carbonell and Brown 1998; Lappin and McCord 1990; Lappin and Leass 1994; Ferrandez et al. 1998). These approaches require a full parse or partial parse of the input sentence. Our approach is a sequential, rule-based, domain-independent procedure, which has been implemented and integrated into the KANTOO English- Spanish MT system. The success rate for anaphora resolution is 89.6% on our training corpus and 87.5% on held-out data. The accuracy rate for translation of pronouns to Spanish is 97.2% on the held-out data. Encoding the world knowledge and deep inference mechanisms required for selecting the right antecedent is a bottleneck in reaching 100% coverage in unrestricted texts for both “knowledge-poor” approaches (Dagan and Itai 1990; Kennedy and Boguraev 1996; Mitkov 1998; Nasukawa 1994) and linguistic approaches. Nasukawa pointed out the difficulty of achieving a success rate of over 90% without this type of knowledge, and our results support this claim. Our algorithm is syntax-based, utilizing the f-structure that results from a full parse using an analysis grammar. The algorithm applies a set of well-known heuristics (constraints and preferences) used in “knowledge-poor” systems. However, it differs from previous approaches in that it does not calculate weights for the heuristics in order to choose the right antecedent; rather it applies heuristics in a sequential manner until one candidate antecedent remains. Since our evaluation indicates performance comparable to that of score-based, knowledge-poor systems, it can be inferred that adding more linguistic knowledge reduces the need for scoring procedures to prune incorrect antecedents. If necessary, semantic knowledge can …

Content truncated to 500 words