Generalizing Sub-sentential Paraphrase Acquisition across Original Signal Type of Text Pairs

Proceedings of the 2012 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning, pages 721–731, Jeju Island, Korea, 12–14 July 2012. c⃝2012 Association for Computational Linguistics Generalizing Sub-sentential Paraphrase Acquisition across Original Signal Type of Text Pairs Aur´elien Max Houda Bouamor LIMSI-CNRS & Univ. Paris Sud Orsay, France firstname.lastname@limsi.fr Anne Vilnat Abstract This paper describes a study on the impact of the original signal (text, speech, visual scene, event) of a text pair on the task of both man- ual and automatic sub-sentential paraphrase acquisition. A corpus of 2,500 annotated sen- tences in English and French is described, and performance on this corpus is reported for an efficient system combination exploiting a large set of features for paraphrase recogni- tion. A detailed quantified typology of sub- sentential paraphrases found in our corpus types is given. 1 Introduction Sub-sentential paraphrases can be acquired from text pairs expressing the same meaning (Madnani and Dorr, 2010). If the semantic similarity of a text pair has a direct impact on the quality of the ac- quired paraphrases, it has, to our knowledge, never been shown what impact the type of original sig- nal has on paraphrase acquisition. In this work, we consider four types of corpora, which we think are representative of the main types of original semantic signals: text pairs (roughly, sentences) originating a) from independent translations of a text (TEXT), b) from independent translations of a speech (SPEECH), c) from independent descriptions of a visual scene (SCENE), and d) from independent descriptions of some event (EVENT). We will report the results of experiments on sub-sentential para- phrase acquisition on all these corpus types in two languages, English and French, and provide some answers to the following questions: What types of paraphrases can be found by human annotators, with what confidence and in which quantities? How well can representative paraphrase acquisition systems perform on each corpus type, and how performance can be improved through combination? On what corpus types can performance be improved by using training material from other corpus types? Our ex- perimental results will provide several indications of the differences and complementarities of the corpus types under study, and will notably show that perfor- mance on the most readily available corpus type can be improved by using training data from the set of all other corpus types. We will first describe the building procedures and characteristics of our corpora (section 2), and then describe our experimental settings for evalu- ating paraphrase acquisition (section 3.1). Our ex- periments will first consist of the description (sec- tion 3.2) and evaluation (section 3.3) of a system combination on each corpus type and then of our system provided with additional training data from the other corpus types (section 3.4). We will finally briefly review related work (section 4) and discuss our main findings and future work (section 5). 2 Collection of sentence pair corpora In this study, we will focus on paraphrase acquisition from related sentence pairs characteristic of 4 …

Content truncated to 500 words