Development of a paraphrase corpus for Turkish short text similarity studies
Project no. 538 · Status: Kapanmış · 2015
Summary
Deciding whether given two sentences are paraphrases of each other is not a deterministic problem. Different people may comprehend the same sentence in different ways. In natural language understanding, the possibility of representing the same meaning with different words and structures and lack of sufficient information in short texts makes the problem of measuring similarity between two sentences harder. Automatically labeling the resemblance of two sentences in Turkish is not a sufficiently studied problem. Automatic measurement of short text similarities have found many applications in the field of Information Retrieval, IR, (Deerwester et.al.,1990; Park et.al, 2005), text summarisation (Erkan and Radev 2004; Lin and Hovy, 2003), assessment of machine translation, MT, (Liu and Zong 2004; Papineni et.al., 2002), assessment of question answering, QA, (Rehder et.al., 1998), disambiguation of word senses, WS, (Lesk, 1986; Schutze, 1998), text categorization (Ko et.al, 2004; Liu and Guo, 2005). Corpus is an important foundation for natural language processing studies. The main goal of this project is to develop a golden standard paraphrase Turkish corpus that will be a useful resource for problems in the fields of IR, NLP and NL understanding that needs the measure of text similarity. To achieve this goal, candidate sentence pairs are automatically extracted from different sources and labeled by human experts as paraphrase or not. Inter-rater agreement assessment analysis is made to assure the consistency of labeling. Paraphrase corpus, the main output of the project, consists of 6510 sentence pairs of which 4421 pairs are labelled as paraphrase and 2089 pairs are labelled as nonparaphrase.