Recognizing Textual Entailment And Its Application In Question Answering

Textual Entailment is one of the most challenging research areas in Natural Language Processing(NLP), and plays an increasingly important role in many NLP applications. As a general framework for textual inference, textual entailment provides a platformon which various kinds of semantic representation, knowledge acquisition and inference approaches can be integrated. Hence textual entailment will have wide application in prospect.This thesis involves two strategies:one is classification strategy; the other is transformation strategy. In classification strategy, Recognizing Textual Entailment (RTE) is formed as a bi-categorization or tri-categorization problem. While in transformation strategy, RTE alternates to estimate whether two text fragments can be derived into an equivalent form by transformation. Thus our work in this paper proceeds around these two strategies. As there is no training data in classification strategy, we propose a Semi-Supervised Learning(SSL) approach, namely Co-training, to recognize entailment relations in texts; As to the problem that semantically complex relation can not be represented using entailed rules in current transformation strategy, we propose Event Model-based approach for recognizing textual entailment. Then we adopt two RTE approaches mentioned above to improve the performance of Question Answering.The primary research and contents can be summarized as follows:1. A huge number of language phenomena abound in both entailment category and non-entailment category. So the entailment degree between two similar text fragments is unstable, which results in the low performance of classification. In order to keep a good and stable performance in terms of inadequate training data, we propose an approach based on Co-training for SSL of RTE. To this end, two views in Co-training are given:one is from the perspective of the rewriter; the other, the assessor. The former stands for a rewriter’s viewpoints in surveying the data and considers the entailment relation as the changes of the syntactic/semantic structures. And the latter view assumes an assessor whose judgment is based on plat similarity features, e.g., lexical overlap or string similarities. By comparing with the supervised classification methods, we find that the performance of the semi-supervised classifier based on Co-training improves remarkably when the labeled data is 30% to 70% of the origin training data.2. We build a deep semantic representation method by using Event Model. Essentially, Event Model is a graph, in which the vertexes are events, and the edges are relations between two events. This type of relation can be a binary relation, namely entailment or no entailment, or a logical relation of discourse. Event Model overcomes the problem of traditional text representation and has the ability of representing deep semantic relations, which can infer the entailment relations across the discourses. In addition, Event Model is robust. By improving the performances of the logical relation recognition and the disambiguation of entailment rules, our RTE system based on Event Model achieves a better performance in comparison with the semi-supervised RTE system implemented using Co-training.3. We propose a RTE approach by using Event Model. Our approach investigates entailment relations between two text fragments in two aspects:the one is the entailment relation between the two fragments, including equivalent and inequivalent relation; the other is the logical relation in each text, i.e., causation and temporal relation. An entailment relation exists between the text and the hypothesis iff the events in hypothesis are entailed by the corresponding events in text, and the logical relations between the events in hypothesis are coherent with the logical relations between the corresponding events in text. To this end, we propose the expanding algorithms, aiming at expands events and the entailment and logical relations among them. Finally, an estimate measure is given for entailment confidence.4. The two RTE approaches proposed are adopted into Question Answering system. The semi-supervised RTE system is integrated into document and passage retrieval as a coarse-gained recognition method, and the Event Model-based RTE system is adopted into question analysis and answer extraction for a fine-gained recognition method. The experiments show that the performance of Question Answering system improves greatly in comparison with the baseline system.