Multi-Task Learning for Coherence Modeling

Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 629–639 Florence, Italy, July 28 - August 2, 2019. c⃝2019 Association for Computational Linguistics 629 Multi-Task Learning for Coherence Modeling Youmna Farag Helen Yannakoudakis Department of Computer Science and Technology The ALTA Institute University of Cambridge United Kingdom {youmna.farag,helen.yannakoudakis}@cl.cam.ac.uk Abstract We address the task of assessing discourse co- herence, an aspect of text quality that is essen- tial for many NLP tasks, such as summariza- tion and language assessment. We propose a hierarchical neural network trained in a multi- task fashion that learns to predict a document- level coherence score (at the network’s top lay- ers) along with word-level grammatical roles (at the bottom layers), taking advantage of in- ductive transfer between the two tasks. We as- sess the extent to which our framework gen- eralizes to different domains and prediction tasks, and demonstrate its effectiveness not only on standard binary evaluation coherence tasks, but also on real-world tasks involving the prediction of varying degrees of coherence, achieving a new state of the art. 1 Introduction Discourse coherence refers to the way textual units relate to one another and form a coher- ent whole. Coherence is an important aspect of text quality and therefore its modeling is essen- tial in many NLP applications, including sum- marization (Barzilay et al., 2002; Parveen et al., 2016), question-answering (Verberne et al., 2007), question generation (Desai et al., 2018), and lan- guage assessment (Burstein et al., 2010; Soma- sundaran et al., 2014; Farag et al., 2018). large body of work has investigated models for theA assessment of inter-sentential coherence, that is, assessment in terms of transitions between adja- cent sentences (Barzilay and Lapata, 2008; Yan- nakoudakis and Briscoe, 2012; Guinaudeau and Strube, 2013; Tien Nguyen and Joty, 2017; Joty et al., 2018). The properties of text that re- sult in inter-sentential connectedness have been translated into a number of computational mod- els – some of the most prominent ones include the entity-based approaches, inspired by Center- ing Theory (Grosz et al., 1995) and proposed in the pioneering work of Barzilay and Lapata (2005, 2008). Such approaches model local coherence in terms of entity transitions between adjacent sen- tences, where entities are represented by their syn- tactic role in the sentence (e.g., subject, object). Current state-of-the-art deep learning adapta- tions of the entity-based framework involve the use of Convolutional Neural Networks (CNNs) over an entity-based representation of text to dis- criminate between a coherent document and its incoherent variants containing a random reorder- ing of the document’s sentences (Tien Nguyen and Joty, 2017); as well as lexicalized counterparts of such models that further incorporate lexical infor- mation regarding the entities, thereby distinguish- ing between different entities (Joty et al., 2018). In contrast to existing approaches, we propose a more generalized framework that allows neural models to encode information about the types of grammatical roles all words in a sentence partic- ipate in, rather than focusing only on the roles of entities within …

Content truncated to 500 words