Error correction using utterance disambiguation techniques

Proceedings of the 2006 Australasian Language Technology Workshop (ALTW2006), pages 123–130. Error correction using utterance disambiguation techniques Peter Vlugter and Edwin van der Ham and Alistair Knott Dept of Computer Science University of Otago Abstract This paper describes a mechanism for identifying errors made by a student dur- ing a computer-aided language learning dialogue. The mechanism generates a set of ‘perturbations’ of the student’s origi- nal typed utterance, each of which em- bodies a hypothesis about an error made by the student. Perturbations are then passed through the system’s ordinary ut- terance interpretation pipeline, along with the student’s original utterance. An utter- ance disambiguation algorithm selects the best interpretation, performing error cor- rection as a side-effect. 1 Introduction The process of identifying and correcting the er- rors in an input utterance has been extensively studied. Kukich (1992) discusses three progres- sively more difficult tasks. The first task is the identification of nonwords in the utterance. nonword is by definition a word which is not partA of the language, and which therefore must must have been misspelled or mistyped. The difficulty in identifying nonwords is due to the impossibility of assembling a list of all the actual words in a lan- guage; some classes of actual words (in particular proper names) are essentially unbounded. So the system should have a reliable way of identifying when an unknown word is likely to belong to such a class. The second task is to suggest corrections for in- dividual nonwords, based purely on their similar- ity to existing words, and a model of the likelihood of different sorts of errors. This task is performed quite well by the current generation of spellcheck- ers, and is to some extent a solved problem. The third task is to detect and correct valid word errors—that is, errors which have resulted in words (for instance their misspelled as there). This task requires the use of context: the error can only be detected by identifying that a word is out of place in its current context. Probabilis- tic language models which estimate the likelihood of words based on their neighbouring words have been used quite successfully to identify valid word errors (see e.g. Golding and Schabes (1996); Mangu and Brill (1997); Brill and Moore (2000)); however, there are situations where these tech- niques are not able to identify the presence of er- rors, and a richer model of context is needed, mak- ing reference to syntax, semantics or pragmatics. In summary, two of the outstanding problems in automated error correction are identifying proper names and detecting and correcting errors which require a sophisticated model of context. In this paper, we consider a domain where these two problems arise with particular force: computer-aided language learning dialogues (or CALL dialogues. In this domain, the system plays the role of a language tutor, and the user is a student learning a target language: the student en- gages with the system in a dialogue on some pre- set topic. One of the system’s …

Content truncated to 500 words