Discovering Features Towards Recognising Textual Entailment

Recognising Textual Entailment (RTE) has the objective of judging whether one piece of short text, normally a single sentence, follows from another piece of longer text. Such a task has been the focus of three consecutive challenges since 2005. Current approaches have explored the more sophisticated deep Natural Language Processing (NLP) techniques. However, the capability of shallow features are yet to be fully exploited. Furthermore, since the data for the task are extracted from different NLP categories, namely Question Answering (QA), Text Summarisation (SUM), Information Retrieval (IR) and Information Extraction (IE), further data/feature analysis can be done on these categories individually to search for differences between these categories in terms of Recognising Textual Entailment. This thesis sets to analyse the RTE datasets for features that can be used in classification, to examine the effectiveness of using only shallow NLP techniques and to investigate whether data from different NLP categories possess different characteristics. The practicality of the shallow features and the performance of the RTE system are analysed in several experiments. The results of the experiments indicate that shallow features are candidates for feature-based RTE systems though further improvements require the extraction of deeper features. Moreover, a system that uses a different set of features on each individual NLP category has the potential of beating a single classifier that is trained on all the training data and all the features in accuracy. Future work include improving the current matching and alignment techniques, better relation detection and adding in temporal normalisation.