Reddit Temporal N-gram Corpus and its Applications on Paraphrase and Semantic Similarity in Social Media using a Topic-based Latent Semantic Analysis

Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers, pages 3553–3564, Osaka, Japan, December 11-17 2016. Reddit Temporal N-gram Corpus and its Applications on Paraphrase and Semantic Similarity in Social Media using a Topic-based Latent Semantic Analysis Anh Dang1, Abidalrahman Moh’d1, Aminul Islam2, Rosane Minghim3, Michael Smit4, and Evangelos Milios1 1{anh,amohd,eem}@cs.dal.ca, Dalhousie University, 6050 University Avenue, Halifax, NS, Canada B3H 4R2 2aminul@louisiana.edu, School of Computing and Informatics, University of Louisiana at Lafayette Lafayette, LA, USA 70503 3rminghim@icmc.usp.br, University of S˜ao Paulo-USP, ICMC, S˜ao Carlos, Brazil 4mike.smit@dal.ca, School of Information Management, Dalhousie University, 6100 University, Halifax, NS, Canada B3H 4R2 Abstract This paper introduces a new large-scale n-gram corpus that is created specifically from social media text. Two distinguishing characteristics of this corpus are its monthly temporal attribute and that it is created from 1.65 billion comments of user-generated text in Reddit. The usefulness of this corpus is exemplified and evaluated by a novel Topic-based Latent Semantic Analysis (TLSA) algorithm. The experimental results show that unsupervised TLSA outperforms all the state-of-the-art unsupervised and semi-supervised methods in SEMEVAL 2015: paraphrase and semantic similarity in Twitter tasks. 1 Introduction A word n-gram is a continuous sequence of n words from a corpus of texts or speech. Word n-gram language models are widely used in Natural Language Processing (NLP), such as speech recognition, machine translation, and information retrieval. The effectiveness of a word n-gram language model is highly dependent on the size and coverage of its training corpus (Clarke et al., 2002). A simple algorithm can outperform a more complicated algorithm if it uses a larger corpus (Norvig, 2008). Many large- scale corpora (Brants and Franz, 2006; Baroni et al., 2009; Wang et al., 2010b) based on web contents have been created for this purpose. As the use of social media is increasing, Online Social Networks (OSNs) have become a norm to spreading news, rumours, and social events (Kwak et al., 2010). This growing usage of social media has created both challenges and opportunities. One major challenge is that social media data is intrinsically short and noisy. A study by (Wang et al., 2010a) revealed that different text corpora have significantly different properties and lead to varying performance in many NLP applications. More importantly, we observe that there is no existing large-scale n-gram corpus that is created specifically from social media text. This has motivated us to create an n-gram corpus that is derived from 1.65 billion comments in the Reddit corpus (Baumgartner, July 2015) and make it available to the research community. There are two main features of this corpus that do not exist in the available large-scale corpora in the literature: monthly time-varying (temporal) and purely social media text. This corpus will allow researchers to analyze and make sense of massive social network text, such as finding corresponding terms across time (Zhang et al., 2015) and improving named entity recognition in tweets (Li and Liu, 2015). Moreover, a cloud-based visualization interface is implemented to allow end users to …

Content truncated to 500 words