Novelty Goes Deep. A Deep Neural Solution To Document Level Novelty Detection

Proceedings of the 27th International Conference on Computational Linguistics, pages 2802–2813 Santa Fe, New Mexico, USA, August 20-26, 2018. 2802 Novelty Goes Deep. A Deep Neural Solution To Document Level Novelty Detection Tirthankar Ghosal†, Vignesh Edithal†, Asif Ekbal†, Pushpak Bhattacharyya†, George Tsatsaronis‡, Srinivasa Satya Sameer Kumar Chivukula‡ †Indian Institute of Technology Patna, Bihar, India ‡Elsevier, Amsterdam, Netherlands †(tirthankar.pcs16, edithal.cs14, asif, pb)@iitp.ac.in ‡(g.tsatsaronis, s.chivukula)@elsevier.com Abstract The rapid growth of documents across the web has necessitated finding means of discarding re- dundant documents and retaining novel ones. Capturing redundancy is challenging as it may involve investigating at a deep semantic level. Techniques for detecting such semantic redun- dancy at the document level are scarce. In this work we propose a deep Convolutional Neural Network (CNN) based model to classify a document as novel or redundant with respect to a set of relevant documents already seen by the system. The system is simple and does not require man- ual feature engineering. Our novel scheme encodes relevant and relative information from both source and target texts to generate an intermediate representation for which we coin the name Relative Document Vector (RDV). The proposed method outperforms the existing benchmark on two document-level novelty detection datasets by a margin of ∼5% in terms of accuracy. We further demonstrate the effectiveness of our approach on a standard paraphrase detection dataset where the paraphrased passages closely resembles semantically redundant documents. 1 Introduction Document-level novelty detection implies the categorization of a document as novel if it contains suffi- cient new information with respect to whatever relevant is previously known or seen. Existing literature on novelty detection from texts mostly followed the detection of novelty at the sentence level, i.e., to extract the sentences that carry new information with respect to some relevant source texts (c.f. Section 2). The current work is aimed at automatically classifying an incoming document as novel or non-novel on the basis of documents already seen by the system. We view the problem as a two-class classification problem in machine learning with the judgment that whether an incoming document bears sufficiently new information to be labeled as novel with respect to a set of source documents. The source document set could be seen as the memory of the reader which stores known information. Document-level novelty detection is a crucial problem and finds application in diverse domains of language processing such as information retrieval, document summarization, predicting impact of scholarly articles, etc. It is a com- plex problem that comprehends lexical, syntactic, semantic and pragmatic levels of texts in conjunction with certain characteristics like relevance, diversity, relativity, and temporality (Ghosal et al., 2018). Lit- erature methods for novelty detection based on lexical similarity, divergence features, and information retrieval measures, although proved effective for sentence level but could not address the document-level comprehension needs. Our understanding of the problem led us to an assertion that a deep neural network might be able to extract the text subtleties involved in understanding a document’s novelty. To the best of our …

Content truncated to 500 words