Semantic Novelty Detection and Characterization in Factual Text Involving Named Entities

Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 9225 - 9252 December 7-11, 2022 ©2022 Association for Computational Linguistics Semantic Novelty Detection and Characterization in Factual Text Involving Named Entities Nianzu Ma†, Sahisnu Mazumder‡, Alexander Politowicz†, Bing Liu†, Eric Robertson♮, Scott Grigsby♮ †Department of Computer Science, University of Illinois at Chicago, USA ‡Intel Labs, USA ♮PAR Government Systems Corporation, USA jingyima005@gmail.com, sahisnumazumder@gmail.com, politow2@uic.edu liub@uic.edu, {eric_robertson,Scott_Grigsby}@partech.com Abstract Much of the existing work on text novelty de- tection has been studied at the topic level, i.e., identifying whether the topic of a document or a sentence is novel or not. Little work has been done at the fine-grained semantic level (or contextual level). For example, given that we know Elon Musk is the CEO of a technology company, the sentence “Elon Musk acted in the sitcom The Big Bang Theory” is novel and sur- prising because normally a CEO would not be an actor. Existing topic-based novelty detection methods work poorly on this problem because they do not perform semantic reasoning involv- ing relations between named entities in the text and their background knowledge. This paper proposes an effective model (called PAT-SND) to solve the problem, which can also charac- terize the novelty. An annotated dataset is also created. Evaluation shows that PAT-SND out- performs 10 baselines by large margins. 1 Introduction Novelty/anomaly detection has been an active re- search area for decades (Grubbs, 1969; Chalapathy et al., 2018; Pang et al., 2021). Recently, it has re- ceived increased attention in NLP. Broadly speak- ing, there are two main types: (1) Topic-based novelty detection, which classifies a given text to a training/known class (topic) or reject/detect it as belonging to some unknown classes (Fei and Liu, 2016; Shu et al., 2017; Lin and Xu, 2019; Zheng et al., 2020); and (2) Semantic novelty de- tection, which determines whether a given text represents a semantically/contextually novel phe- nomenon. For example, Ma et al. (2021) studied a semantic novelty detection problem - detecting semantically novel scene descriptions (e.g., “A per- son walks a chicken in the park” is a novel scene, whereas “A person walks a dog in the park” is normal one). This task is more fine-grained and requires factual reasoning over text as compared to that of (1), which has been studied extensively. (2) d1 : “The Big Bang Theory is an American televi- sion sitcom, filmed in front of a live audience, stars Johnny Galecki et al." normal d2 : “Elon Musk’s performance as a dishwasher in a restaurant in season 9, episode 9 of the The Big Bang Theory is quite interesting to his fans." novel Figure 1: Examples of semantic novelty detection in factual texts involving named entities. has only been introduced recently and is the focus of this paper. This paper proposes a new semantic novelty de- tection task: given a factual text d containing two named entities1, we want to classify whether d rep- resents a semantically novel fact or a normal …

Content truncated to 500 words