Multi-Level Alignments As An Extensible Representation Basis for Textual Entailment Algorithms

Proceedings of the Fourth Joint Conference on Lexical and Computational Semantics (*SEM 2015), pages 193–198, Denver, Colorado, June 4–5, 2015. Multi-Level Alignments As An Extensible Representation Basis for Textual Entailment Algorithms Tae-Gil Noh (noh@cl.uni-heidelberg.de)1, Sebastian Pad´o (pado@ims.uni-stuttgart.de)2, Vered Shwartz (vered1986@gmail.com)3, Ido Dagan (dagan@cs.biu.ac.il)3, Vivi Nastase (nastase@fbk.eu)4, Kathrin Eichler (kathrin.eichler@dfki.de)5, Lili Kotlerman (lili.dav@gmail.com)3, and Meni Adler (meni.adler@gmail.com)3 1Institute of Computational Linguistics, Heidelberg University, Germany 2Institute of Natural Language Processing, Stuttgart University, Germany 3Department of Computer Science, Bar-Ilan University, Israel 4Human Language Technologies, Fondazione Bruno Kessler, Italy 5Language Technology Lab, DFKI GmbH, Germany Abstract A major problem in research on Textual Entail- ment (TE) is the high implementation effort for TE systems. Recently, interoperable standards for annotation and preprocessing have been proposed. In contrast, the algorithmic level remains unstandardized, which makes compo- nent re-use in this area very difficult in prac- tice. In this paper, we introduce multi-level alignments as a central, powerful representa- tion for TE algorithms that encourages modu- lar, reusable, multilingual algorithm develop- ment. We demonstrate that a pilot open-source implementation of multi-level alignment with minimal features competes with state-of-the- art open-source TE engines in three languages. 1 Introduction A key challenge of Natural Language Processing is to determine what conclusions can be drawn from a natural language text, a task known as Textual Entail- ment (TE, Dagan and Glickman 2004). The ability to recognize TE helps dealing with surface variability in tasks like Question Answering (Harabagiu and Hickl, 2006), Intelligent Tutoring (Nielsen et al., 2009), or Text Exploration (Berant et al., 2012). Open source implementations a number of TE algorithms have be- come available over the last years, including BIUTEE (Stern and Dagan, 2012) and EDITS (Kouylekov and Negri, 2010), which has made it much easier for end users to utilize TE engines. At the same time, the situation is still more difficult for researchers and developers. Even though recently a common platform for TE has been proposed (Pad´o et al., 2015) that standardizes important aspects like annotation types, preprocessing, and knowledge re- sources, it largely ignores the algorithmic level. In fact, TE algorithms themselves are generally not de- signed to be extensible or interoperable. Therefore, changes to the algorithms – like adding support for a new language or for new analysis aspect – are often very involved, if not impossible. This often forces the next generation of TE researchers to develop and implement their own core algorithms from scratch. In this paper, we address this problem by propos- ing a schema for TE algorithms that revolves around a central representation layer called multi-level align- ment geared towards encoding the relevant informa- tion for deciding entailment. The use of multi-level alignments encourages a modular, extensible devel- opment of TE algorithms that can be partitioned into “alignment producers” and “alignment consumers”. This enables for future researchers and developers to change analysis components or add new ones in a straightforward manner. We also present evaluation results for a very simple TE algorithm based on multi-level alignments for English, German and …

Content truncated to 500 words