Reduction of Search Space to Annotate Monolingual Corpora

Proceedings of the 5th International Joint Conference on Natural Language Processing, pages 1457–1461, Chiang Mai, Thailand, November 8 – 13, 2011. c⃝2011 AFNLP Reduction of Search Space to Annotate Monolingual Corpora Prajol Shrestha Christine Jacquin Laboratore d’Informatique de Nantes-Atlantique (LINA) Universit´e de Nantes 44322 Nantes Cedex 3, France {prajol.shrestha;christine.jacquin;beatrice.daille}@univ-nantes.fr Beatrice Daille Abstract Monolingual corpora which are aligned with similar text segments (paragraphs, sentences, etc.) are used to build and test a wide range of natural language processing applications. The drawback wanting to use them is the lack of publicly available an- notated corpora which obligates people to make one themselves. The annotation pro- cess is a time consuming and costly task. This paper describes a new corpus-based measure to significantly reduce the search space for a faster and easier manual an- notation process for monolingual corpora. This measure can be used in making align- ments on different types of text segments. The performance of this measure is eval- uated on a manually annotated paragraph corpus, whose alignments are freely avail- able, with promising results. 1 Introduction In the field of Natural Language Processing (NLP), annotated monolingual corpora are used to build and test a wide range of applications such as information retrieval, summarization, plagiarism detection, dictionary building and so on. With a number of applications to be built, the field of NLP requires a wide range of monolingual cor- pus with different annotations on different level of text segments. Our focus is on the fast and easy way of aligning short text segments based on sim- ilarity. These annotations are usually done man- ually by annotators as done by Hatzivassiloglou et al. (1999) where a corpus with alignments be- tween similar short texts are created by two or more annotators who look at each possible short text pair independently and analyse them to make a decision on whether each pair should be aligned as similar or not. Finally, the annotators discuss the disagreements between their annotations and come to an agreement with reasoning. A cor- pus containing n number of short texts will gener- ate n(n−1) 2 number of short text pairs for compar- ing similarities which becomes a tedious and time consuming task even if a corpus contains a few hundred of short texts. For example, the corpus we use consists of 239 paragraphs, explained in section 3.1, generating a total of 28,441 text pairs to compare. There are few publicly available annotated cor- pus, some are manually annotated like the TDT corpus1 for topic detection and tracking and the METER corpus (Gaizauskas et al., 2001) for de- tection of text reuse and some are automatically annotated like the PAN-PC-10 (Barr´on-Cede˜no et al., 2010) for plagiarism detection and the MSRPC (Dolan and Brockett, 2005) for paraphrase detec- tion. Annotating a corpus automatically is easier and faster than manual annotation but they have a major limitation which allows the corpus to in- clude only a subset of the problem which prevents the corpus to represent many of the naturally oc- …

Content truncated to 500 words