Alexander M. Franz

Spoken language recognition in an office management domain

The authors highlight needs related to a voice interface and describe the implementation of a general-purpose spoken language interface, the Carnegie Mellon Spoken Language Shell (CM-SLS). CM-SLS provides voice interface services to different applications running on the same computer. CM-SLS was used to build the Office Manager, a collection of applications that includes an appointment calendar, a personal database, voice mail, and a calculator. The performance of several system components is described.<<ETX>>

Coping with ambiguity in a large-scale machine translation system

In an interlingual knowledge-based machine translation system, ambignuity arises when the source language analyzer produces more than one interlingua expression for a source sentence. This can have a negative impact on translation quality, since a target sentence may be produced from an unintended meaning. In this paper we describe the methods used in the KANT machine translation system to reduce or eliminate ambiguity in a large-scale application domain. We also test these methods on a large corpus of test sentences, in order to illustrate how the different disambiguation methods reduce the average number of parses per sentence.

An integrated architecture for example-based machine translation

This paper describes a machine translation architecture that integrates the use of examples for flexible, idiomatic translations with the use of linguistic rules for broad coverage and grammatical accuracy. We have implemented a prototype for English-to-Japanese translation, and our evaluation shows that the system has good translation quality, and only requires reasonable computational resources.

Learning PP attachment from corpus statistics

One of the main problems in natural language analysis is the resolution of structural ambiguity. Prepositional Phrase (PP) attachment ambiguity is a particularly difficult case. We describe a robust PP disambiguation procedure that learns from a text corpus. The method is based on a loglinear model, a type of statistical model that is able to account for combinations of multiple categorial features. A series of experiments that compare the loglinear method against other strategies are described. For the difficult case of three possible attachment sites, the loglinear method predicts PP attachment with significantly higher accuracy than a simpler procedure that uses lexical association strengths. At the same time, on general newswire text, the accuracy of the statistical method remains 10% below the performance of human experts. This suggests a limit on what can be learned automatically from text, and points to the need to combine machine learning with human expertise.

A Model for Part-of-Speech Prediction

Robust natural language analysis systems must be able to handle words that are not in the lexicon. This paper describes a statistical model that predicts the most likely Parts-of-Speech for previously unseen words. The method uses a loglinear model to combine a number of orthographic and morphological features, and returns a probability distribution over the open word classes. The model is combined with a stochastic Part-of-Speech tagger to provide a model of context. Empirical evaluation shows that this results in significant gains in Part-of-Speech prediction accuracy over simpler methods.