Context-Sensitive Natural Language Inferences

Supported by IBM

Period: 1999

The project studies a machine learning centered approach to data-intensive and computing-intensive processing for intelligent context-sensitive human-machine interfaces. The future of intelligent human-machine interaction is in the ability to perform context-sensitive inferences. Most of the difficulties encountered in trying to support free-style queries, text and speech interpretation and other HCI problems involve problems of context-sensitive ambiguities. These problems occur at different levels of processing, from word-sense disambiguation (e.g, the use of the word "plant" in different contexts), to phrase interpretation (e.g., "move the car with the flag"). Additional crucial factors are those of incomplete utterances that humans easily complete based on the context, and human errors, typing or cognitive, that result in valid words and may mislead the interpretation if not corrected based on context-sensitive information. (E.g. typing "our" instead of "out", "now" instead of "know"). All these interact to make intelligent human-machine interfaces a difficult problem. Context sensitive inferences are knowledge intensive tasks that are hard to resolve without a significant learning (knowledge acquisition) component. The learning approach developed in the Cognitive Computation Lab. directly addresses the main challenge in building a realistic system with context-sensitive inference capabilities - that of scalability. It is centered around a novel learning architecture that is specifically tailored towards large scale processes, in terms of both data and computation. The approach developed can be applied to support a variety of inferences of the sort required in intelligent human-machine interactions. One of the available demonstrations, a wide coverage and accurate context-sensitive text correction is available at EOH (Context-Sensitive Text Correction is the task of fixing spelling errors that happen to result in valid words, such as substituting "to" for "too", "casual" for "causal" or simple word usage errors like using "amount" instead of "number". ) In it's current form the system has been trained to represent (and correct mistakes for) about 1000 words. Training was done by having the system read a collection of over 100,000 (un-annotated) sentences containing over 2 million words from the Wall Street Journal. This process takes on our current machines only a few minutes, but require a very large memory.