Using Inconsistency Detection to Overcome Structural Ambiguity in Language Learning

This paper proposes the Inconsistency Detection Learner (IDL), an algorithm for language acquisition intended to address the problem of structural ambiguity. An overt, acoustically audible form is structurally ambiguous if different languages admitting the overt form would assign it different linguistic structural analyses. Because the learner has to be capable of learning any possible human language, and because the learner is dependent on overt data to determine what the target language is, the learner must be capable ultimately of inferring which analysis of an ambiguous overt form is correct by reference to other overt data of the language. IDL does this in a particularly direct way, by attempting to construct hypothesis grammars for combinations of interpretations of the overt forms, and discarding those combinations that are shown to be inconsistent. A specific implementation of IDL is given, based on Optimality Theory. Results are presented from a computational experiment in which this implementation of IDL was applied to all possible languages predicted by an Optimality theoretic system of metrical stress grammars. The experimental results show that this learning algorithm learns quite efficiently for languages from this system, completely avoiding the potential combinatoric growth in combinations of interpretations, and suggesting that this approach may play an important role in the acquisition mechanisms of human learners. Using Inconsistency Detection to Overcome Structural Ambiguity in Language Learning Bruce Tesar Department of Linguistics Rutgers Center for Cognitive Science Rutgers University, New Brunswick 9/12/00 1. Structural Ambiguity in Language Learning 1.1. Mutual Entanglement A central challenge of learning natural languages is that of contending with input data that are structurally ambiguous. The portion of an utterance that is directly perceivable by the learner, labeled here the overt form, is structurally ambiguous if there is more than one complete structural description that may be assigned to it. The situation we are concerned with in this paper is that where the different structural descriptions are grammatical in different languages. Ambiguity within a language, where the same overt form can be assigned more than one analysis by a single language, is not of direct concern here. Structural ambiguity can be illustrated with metrical stress theory. For present purposes, assume that a structural description of a word consists of the ordered sequence of syllables of the word, a grouping of syllables into feet, and an assignment of a stress level to each syllable. The overt form corresponding to a structural description is the ordered string of syllables, along with the stress levels of the syllables. An overt form is not itself a structural description; it only contains structures for elements that are presumed to be directly observable when a child hears a word uttered. What is missing from the overt form is the foot structure; the child cannot directly ‘hear’ foot boundaries. An overt form is ambiguous when more than one structural description shares that overt form. We will refer to a full structural description consistent with an overt form as an interpretation of that overt form. An ambiguous overt form has more than one interpretation. A simple example is a three-syllable word with medial main stress: [ σ σ σ ] (we use σ to denote a syllable). This overt form is ambiguous between at least two interpretations, including: * I would like to thank Jason Eisner, Janet Fodor, Brett Hyde, Jacques Mehler, Joe Pater, Alan Prince, Ken Safir, William Sakas, Vieri Samek-Lodovici, Paul Smolensky, the students of the Spring 2000 Rutgers University Learnability and Linguistic Theory seminar, and the audiences at HOT’97, NELS 28, The CUNY Graduate Center, Carnegie Mellon University, the NELS 30 Workshop on Language Learnability, NYU, MIT, Rochester University, Western Michigan University, and SUNY Stony Brook, for useful comments. Alan Prince also provided many useful comments on an earlier draft of this paper. Part of this research was funded by postdoctoral support from the Department of Linguistics, Rutgers University, and the Rutgers Center for Cognitive Science. 1 The stress levels in the overt form are a direct translation of the relative prominence of the syllables as expressed in acoustic, observable properties: duration, pitch, and amplitude. The syllable structure itself is, of course, constructed by the learner, based upon the acoustic signal. Syllable structure construction will simply be assumed for the discussion of stress in this paper, but in general elements of syllable structure can also be subject to cross-linguistic structural ambiguity. 2 Another possible interpretation is one with the stressed syllable as a foot by itself, [ σ ( σ ) σ ]. Such a foot is not unexpected in trochaic, quantity-sensitive languages, when the syllable is heavy.