AAAI 2013 Tutorial on Textual Entailment

Sebastian Pado, Heidelberg University
Rui Wang, German Research Center for Artificial Intelligence

Introduction

The ability to draw inferences is a central part of language understanding. Traditionally, it has predominantly been approached through the development of formal representa- tions with provably complete and correct reasoning mechanisms. Textual Inference is a recent alternative approach that defines inference as a binary relation between two natural language text fragments, avoiding commitment to any specific knowledge repre- sentation, or reasoning mechanism. It hopes to establish a level playing field to compare and combine various approaches to inference and to establishing a task-independent paradigm for applied semantics. The proposed tutorial provides an introduction to Tex- tual Inference, spanning the range from fundamental to applied aspects. It would cover (a) definition and motivation; (b) major relevant linguistic phenomena; (c) methods for acquiring and applying linguistic knowledge for modeling inference; (d), major families 1 of algorithms; and (e) an introduction to practical system building using a new open and modular software platform.

Syllabus

Part 1. Basics: Inference and Textual Inference

  • Discuss the importance of inference for AI and Natural Language Processing;
  • Motivate Textual Inference (TI) as a probabilistic concept of inference on natural language;
  • Show the mapping of major NLP tasks (such as Question Answering and Machine Translation Evaluation) onto Textual Inference;
  • Discuss the relationship between TI and related tasks (e.g., paraphrasing, contradiction recognition);
  • Describe RTE (“Recognizing Textual Entailment”), the main forum for TI evaluation in the NLP community.

Part 2. Classes of Strategies and Learning

  • Characterize the major classes of strategies to approach TI;
  • Cover typical linguistic preprocessing steps;
  • Define representations/data structures typically used;
  • Provide technical detail on state-of-the-art inference procedures and machine learning techniques for the two most practically relevant classes of strategies (transformation-based and classification-based TI).

Part 3. Knowledge Acquisition

  • Discuss the importance of background knowledge in TI;
  • Identify knowledge resources used in current TI systems and their limitations;
  • Provide an overview of (different types of) knowledge acquisition approaches;
  • Define suitable representations and algorithms for using knowledge, including context-sensitive knowledge application.

Part 4. Applications

  • Outline main classes of NLP problems and their mapping into Textual Inference;
  • Go into technical detail for two applications:
  • Machine Translation Evaluation;
  • Hierarchical Information Exploration with Entailment Graphs.

Part 5. Multilingual, Component-based System Building

  • Describe the problems for evaluation and practical application raised by the predominance of research prototype systems in the TI area;
  • Present the component-based platform developed in the EXCITEMENT project (http://www.excitement-project.eu);
  • System demonstration.

Slides for Download

All slides are landscape, printed 2-up on a portrait page (A4).

  1. Slides for Part 1 (2 MB, updated 7/12/13)
  2. Slides for Part 2 (2 MB, updated 7/12/13)
  3. Slides for Part 3 (2 MB, updated 7/12/13)
  4. Slides for Part 4 (2 MB, updated 7/12/13)
  5. Slides for Part 5 (2 MB, updated 7/12/13)
  6. All slides in one file (17 MB, 7/12/13)