Introspective Knowledge Revision in Textual Case-Based Reasoning

The performance of a Textual Case-Based Reasoning system is critically dependent on its underlying model of text similarity, which in turn is dependent on similarity between terms and phrases in the domain. In the absence of human intervention, term similarities are often modelled using co-occurrence statistics, which are fragile unless the corpus is truly representative of the domain. We present the case for introspective revision in TCBR, whereby the system incrementally revises its term similarity knowledge by exploiting conflicts of its representation against an alternate source of knowledge such as category knowledge in classification tasks, or linguistic and background knowledge. The advantage of such revision is that it requires no human intervention. Our experiments on classification knowledge show that revision can lead to substantial gains in classification accuracy, with results competitive to best-in-line text classifiers. We have also presented experimental results over synthetic data to suggest that the idea can be extended to improve case-base alignment in TCBR domains with textual problem and solution descriptions.