Breaking Through the 80% Glass Ceiling: Raising the State of the Art in Word Sense Disambiguation by Incorporating Knowledge Graph Information

Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 2854–2864 July 5 - 10, 2020. c⃝2020 Association for Computational Linguistics 2854 Breaking Through the 80% Glass Ceiling: Raising the State of the Art in Word Sense Disambiguation by Incorporating Knowledge Graph Information Michele Bevilacqua and Roberto Navigli Sapienza NLP Group Department of Computer Science Sapienza University of Rome {bevilacqua,navigli}@di.uniroma1.it Abstract Neural architectures are the current state of the art in Word Sense Disambiguation (WSD). However, they make limited use of the vast amount of relational information encoded in Lexical Knowledge Bases (LKB). We present Enhanced WSD Integrating Synset Embed- dings and Relations (EWISER), a neural su- pervised architecture that is able to tap into this wealth of knowledge by embedding infor- mation from the LKB graph within the neural architecture, and to exploit pretrained synset embeddings, enabling the network to predict synsets that are not in the training set. As a re- sult, we set a new state of the art on almost all the evaluation settings considered, also break- ing through, for the first time, the 80% ceil- ing on the concatenation of all the standard all- words English WSD evaluation benchmarks. On multilingual all-words WSD, we report state-of-the-art results by training on nothing but English. 1 Introduction There is a growing body of research dealing with the integration of prior knowledge into neural net- works for Natural Language Processing (NLP) tasks, be it through pretraining on self-supervised tasks such as language modeling (Peters et al., 2018; Devlin et al., 2019), or through the incorpora- tion of information from knowledge bases (Peters et al., 2019; Logan et al., 2019). In Word Sense Dis- ambiguation (WSD), i.e., the task of associating a word in context with the most appropriate meaning from a finite set of possible choices (Navigli, 2009), the gap between supervision and knowledge (Nav- igli, 2018) has been overcome by several efforts directed at learning effective vector representations (Loureiro and Jorge, 2019; Scarlini et al., 2020) in the same space as contextualized embeddings, and exploring the usage of definitional knowledge in supervised sequence learning neural architectures (Luo et al., 2018; Kumar et al., 2019; Huang et al., 2019). However, the Lexical Knowledge Bases (LKBs) from which such information is retrieved, such as WordNet (Miller, 1995) and BabelNet (Navigli and Ponzetto, 2012), also provide a great wealth of relational knowledge in structured form (i.e., hy- pernymy, meronymy, similarity, etc.), which is of- ten neglected due to the non-trivial integration of data of this kind into neural architectures. Even though such information can, instead, be exploited by knowledge-based WSD algorithms (Agirre and Soroa, 2009; Moro et al., 2014), rivaling super- vised pre-contextualized embedding approaches (Maru et al., 2019), the performances still lag be- hind (Huang et al., 2019; Vial et al., 2019). Building on Extended WSD Integrating Sense Embeddings (EWISE) (Kumar et al., 2019), a neu- ral WSD system incorporating prior knowledge through synset embeddings, we present Enhanced WSD Integrating Synset Embeddings and Relations (EWISER), a …

Content truncated to 500 words