Topic Classification using RNN: A Combined Approach towards Topic Discovery

Abstract

Dr. Yungcheol Byun is a full professor at the Computer Engineering Department (CE) at Jeju National University (http://www.jejunu.ac.kr). His research interests include the areas of Pattern Recognition & Image Processing, Artificial Intelligence & Machine Learning, Pattern-based Security, Home Network and Ubiquitous Computing, u-Healthcare, and RFID & IoT Middleware System. He directs the Machine Laboratory at the CE department. Recently, he studied at University of Florida as a visiting professor from 2012 to 2014. He is currently serving as a director of Information Science Technology Institute, and other academic societies. Outside of his research activities, Dr. Byun has been hosting international conferences including CNSI (Computer, Network, Systems, and Industrial Engineering), ICESI (Electric Vehicle, Smart Grid, and Information Technology), and serving as a conference and workshop chair, program chair, and session chair in various kinds of international conferences and workshops. Dr. Byun was born in Jeju, Korea, and received his Ph.D. and MS from Yonsei University (http://www.yonsei.ac.kr) in 1995 and 2001 respectively, and BS from Jeju National University in 1993. Before joining Jeju National University, he worked as a special lecturer in SAMSUNG Electronics (http://www.samsung.com) in 2000 and 2001. From 2001 to 2003, he was a senior researcher of Electronics and Telecommunications Research Institute (ETRI, https://etri.re.kr/eng/main/main.etri). He was promoted to join Jeju National University as an assistant professor in 2003.

Biography

In natural language processing (NLP), language model is doubtlessly an intrinsic element, as it plays a fundamental role in many conventional NLP tasks, e.g., speech recognition to image captioning etc. Therefore, learning an exceptional language model usually enhance the hidden aspects or metrics; forging its pivotal role in NLP. Language models are gaining popularity as of the abundance of online texts, comments and reviews. Due to the advancement of e-commerce, people do write their reviews about the products they have received. In crowdfunding sites, comments are so critical that negative reviews can damage the reputation of the product’s creator or can affect the buying of others. Life is too fast these days that people find it difficult to go through abundant of text data to take a decision. Therefore, topic discovery is quite valuable in various aspects as of saving time of the user, providing the summary of text in form of discussion topics, and providing contextual information etc. Topic models are being studied for decades and are of fundamental importance as these models act as a tool in order to infer the latent topics and extracting semantic structure of a document. In this speech, we have used the Latent Topic Model (LDA) in order to generate topics for crowdfunding comments. Our proposed model is recurrent neural network (RNN) based language model, which uses the latent topics generated by LDA, is constructed to extract the comprehensive semantic meaning related words in comments. Moreover, this combined approach is of better capability on creating topic clusters then traditional ones, which signifies that blending the information from deep learning and topic modeling is a substantial way to generate an improved understanding of crowdfunding comments.