Recovering From Parser Failures: A Hybrid Statistical/Symbolic Approach

Recovering From Parser Failures: A Hybrid Statistical/Symbolic Approach Carolyn Penstien Ros& Computational Linguistics Program Carnegie Mellon University cprose@Icl, cmu. edu Alex Waibel Dept. of Computer Science Carnegie Mellon University ahw@cs, cmu. edu Abstract W~ describe an implementation of a hybrid statisti- cal/symbolic approach to repairing parser failures in a speech-to-speech translation system. I We describe a rnodale which takes as input a fragmented parse and ree- turas a repaired meaning representation. It negotiates with the speaker about what the complete meaning of the utterance is by generating hypotheses about how to fit. the fragments of the partial parse together into a colwrcnt meaning representation. By drawing upon both statistical and symbolic information, it constrains its rcpair hypotheses to those which are both likely and meaningful. Because it updates its statistical model during use, it improves its performance over time. Introduction Natural language processing of spontaneous speech is particularly difficult because it contains false starts, out of vocabulary words, and ungrammatical constructions. lleca,se of this, it is unreasonable to hope to be able to write a grammar which will cover all of the phenom- ena which a parser is likely to encounter in a practical speech translation system. In this paper we describe an implementation of a hybrid statistical/symbolic ap- proach to recovering from parser failures in the context of a speech-to-speech translation system of significant scope (vocabulary size of 996, word recognition accu- racy 60 %, grammar size on the order of 2000 rules). The domain which the current system focuses on is the scheduling domain where two speakers attempt to set ~i) a meeting.over the phone. Because this is an interlingua-based translation sys- tem, the goal of the analysis stage of the translation pro- ccss is to map the utterance in the source language onto a f<.at.re-structure representation called an interlingua which represents meaning in a language-independent way. (This approach extends to other feature structure based meauistg representations as well.) If the parser 1'l'hc research described in this paper was sponsored by I1.~ Ih'partmeat of the Navy, Office of Naval Reseaxch, grant //. N01)()14-93.1-0806. The ideas described in this paper do .-I..f,ce.~sarily rellcct the position or the policy of the gov- ,.r.m~l~t,, aad ao official elldorsement should be inferred. cannot derive a complete analysis for an utterance, it derives a partial parse by skipping over portions of the utterance in order to find a subset which can parse. It also returns an analysis for the skipped portions which can be used to rebuild the meaning of the input utter- ance. The goal of our repair module is to interactively reconstruct the meaning of the full utterance by gen- erating predictions about the way the fragments can fit together and checking them with the user. In this way it negotiates with the user in order to recover the meaning of the user's utterance. The repair module described in this paper uses both symbolic and statistical information in order to recon- struct the speaker's meaning from the partial analy- sis which …

Content truncated to 500 words