Fabio Massimo Zanzotto


Textual Entailment Recognition for Web-based Question-Answering

Tutorial at Web Intelligence 2012

Instructor: Fabio Massimo Zanzotto


After shortly motivating and introducing the general task and the evaluation model, the main part of the course is a methodological presentation of some of the key computational approaches proposed to the overall task, organized by major components. In the last part of the course, we describe the view of Semantics that RTE suggests and how applications can use RTE systems.
Tutorial's Material


  1. Motivation and Task Definition
    We will describe the applied notion of Textual Entailment as a generic (application independent) semantic inference test, motivate the task and outline the existing evaluation framework and put it forward as a framework for the study of applied semantics.
  2. Knowledge Representation and Preprocessing
    We will discuss the different kinds of abstractions that are being used over the surface representations and the preprocessing tools that are commonly used.
  3. Knowledge Acquisition Methods
    Having a knowledge base of linguistic and world knowledge has been shown to be a critical component of determining textual entailment. We will present some of the work done on the Knowledge Base component, specifically, knowledge acquisition algorithms and some of the existing resources that might be useful for Textual Entailment.
  4. Inference methods
    Textual Entailment is viewed as a global problem that requires incorporating decisions made by multiple different inference components. We review this view of the problem and identify some of the components and computational approaches that were presented in the literature to address it.
  5. Machine Learning Methods
    Machine Learning methods are, naturally, a significant component of all preprocessing tools used in order to generate a more abstract representation of the Textual Entailment input. However, there are several ways in which one can incorporate Machine Learning directly into the decision problem. These methods vary from tuning the weights of different modules taking part in the entailment inference to direct decision trained over knowledge based feature representation of the candidate pair. We will review several methods considered so far for incorporating machine learning into deciding entailment.
  6. Utility for applications
    We demonstrate initial utility of the Textual Entailment setting for applications and propose further ways in which generic entailment models and engines can be utilized by specific applications.