A Spreading Activation Framework for Tracking Conceptual Complexity of Texts

Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 3878–3887 Florence, Italy, July 28 - August 2, 2019. c⃝2019 Association for Computational Linguistics 3878 A Spreading Activation Framework for Tracking Conceptual Complexity of Texts Ioana Hulpus¸1, Sanja ˇStajner2 and Heiner Stuckenschmidt1 1Data and Web Science Group, University of Mannheim, Germany 2 Symanto Research, N¨urnberg, Germany {ioana,heiner}@informatik.uni-mannheim.de sanja.stajner@symanto.net Abstract We propose an unsupervised approach for as- sessing conceptual complexity of texts, based on spreading activation. Using DBpedia knowledge graph as a proxy to long-term memory, mentioned concepts become acti- vated and trigger further activation as the text is sequentially traversed. Drawing inspira- tion from psycholinguistic theories of reading comprehension, we model memory processes such as semantic priming, sentence wrap-up, and forgetting. We show that our models cap- ture various aspects of conceptual text com- plexity and significantly outperform current state of the art. 1 Introduction Reading comprehension has long been linked to processes over semantic memory, such as seman- tic priming through spreading activation (Ander- son, 1981; Collins and Loftus, 1975; Neely, 1991; Gulan and Valerjev, 2010). While psycholinguis- tic literature abounds in research and demonstra- tion of such processes (Just and Carpenter, 1980; Kutas and Hillyard, 1984; Carroll and Slowiaczek, 1986), there is a gap in understanding if they can be modeled in an automated way for capturing the cognitive load required by texts. At the same time, the recent advances in the publication of en- cyclopedic knowledge graphs provide an unprece- dented opportunity for modeling human knowl- edge at scale. We focus on conceptual complexity which, as opposed to lexical and syntactic complexity (Vaj- jala and Meurers, 2014; Ambati et al., 2016), has received very little attention so far. Conceptual complexity accounts for the background knowl- edge necessary to understand mentioned concepts as well as the implicit connections that the reader has to access between the mentioned concepts in order to fully understand a text. It plays an im- portant role in making texts accessible to children, non-native speakers, as well as people with low lit- eracy levels or intellectual disabilities (Arf´e et al., 2017). Apart from being one of the main fac- tors for understanding the story, conceptual com- plexity also influences the readers’ interest in the text: readers who lack relevant background knowl- edge have difficulties in understanding conceptu- ally complex texts (Arf´e et al., 2017; Benjamin, 2012), while high-knowledge readers need some obstacles (more conceptual complexity) to main- tain their interest (Arf´e et al., 2017; Benjamin, 2012; Kalyuga et al., 2003). Therefore, correctly estimating conceptual complexity of a text, and of- fering a reader a text of an appropriate cognitive load, is of utmost importance for: (1) ensuring cor- rect understanding of a text; (2) maintaining the readers’ interest; and (3) promoting deeper-level processing and enhancing the readers knowledge. In this paper, we are building on top of the psy- cholinguistic findings that words are recognized faster if preceded by words related in meaning (se- mantic priming) (Gulan and Valerjev, 2010), and …

Content truncated to 500 words