Information Retrieval from Unstructured Web Text Document Based on Automatic Learning of the Threshold

Abstract

Collocation is defined as a sequence of lexical tokens which habitually co-occur. This type of information is widely used in various applications such as Information Retrieval, document indexing, machine translation, lexicography, etc. Therefore, many techniques are developed for the automatic retrieval of collocations from textual documents. These techniques use statistical measures based on a joint frequency calculation to quantify the connection strength between the tokens of a candidate collocation. The discrimination between relevant and irrelevant collocations is performed using a priori fixed threshold. Generally, the discrimination threshold estimation is performed manually by a domain expert. This supervised estimation is considered as an additional cost which reduces system performance. In this paper, the authors propose a new technique for the threshold automatic learning to retrieve information from web text document. This technique is mainly based on the usual performance evaluation measures (such as ROC and Precision-Recall curves). The results show the ability to automatically estimate a statistical threshold independently of the treated corpus.