Building Semantic Grams of Human Knowledge

Proceedings of the 12th Conference on Language Resources and Evaluation (LREC 2020), pages 2991–3000 Marseille, 11–16 May 2020 c⃝ European Language Resources Association (ELRA), licensed under CC-BY-NC 2991 Building Semantic Grams of Human Knowledge Valentina Leone1, Giovanni Siragusa1, Luigi Di Caro1, Roberto Navigli2 1Computer Science Department, University of Turin, Italy 2 Computer Science Department, Sapienza University of Rome, Italy {leone, siragusa, dicaro}@di.unito.it navigli@di.uniroma1.it Abstract Word senses are typically defined with textual definitions for human consumption and, in computational lexicons, put in context via lexical-semantic relations such as synonymy, antonymy, hypernymy, etc. In this paper we embrace a radically different paradigm that provides a slot-filler structure, called “semagram”, to define the meaning of words in terms of their prototypical semantic information. We propose a semagram-based knowledge model composed of 26 semantic relationships which integrates features from a range of different sources, such as computational lexicons and property norms. We describe an annotation exercise regarding 50 concepts over 10 different categories and put forward different automated approaches for extending the semagram base to thousands of concepts. We finally evaluate the impact of the proposed resource on a semantic similarity task, showing significant improvements over state-of-the-art word embeddings. We release the complete semagram base and other data at http://nlp.uniroma1.it/semagrams. Keywords: Semagrams, word senses, concept representation, lexical semantics 1. Introduction The representation of knowledge is one of the great dreams of Artificial Intelligence. Acquiring and encoding knowl- edge is essential not only to improve historical NLP tasks such as Word Sense Disambiguation and Machine Trans- lation, but also to enable numerous applications (Hovy et al., 2013) like intelligent personal assistants, Question Answering, Information Retrieval, etc. Lexical-semantic knowledge has typically been encoded on a large scale us- ing word senses as meaning units, starting from WordNet (Miller, 1995) and then carrying on with VerbNet (Schuler, 2005), PropBank (Kingsbury and Palmer, 2002) and, more recently, VerbAtlas (Di Fabio et al., 2019). Large reposito- ries of frames have been introduced with FrameNet (Baker et al., 1998) and its counterparts in other languages. How- ever, frames are focused on situation-based representations with semantic roles and commonly involved objects, with- out defining the embodied ontological semantics of single concepts. Word senses, on the other hand, suffer from a lack of ex- plicit common-sense semantic information. Indeed, the well-known fine granularity of word senses in WordNet (Palmer et al., 2007) is due to the lack of a meaning en- coding system capable of managing the representation of concepts in a flexible way. To address this issue, Puste- jovsky introduced an innovative model (Pustejovsky, 1991) based on qualia roles to enable the creation of new mean- ings via semantic slot filling. However, the approach was limited by the number and type of these slots. An interesting and novel extension was presented by Mo- erdijk et al. (2008) with the ANW dictionary and the in- troduction of the concept of semagram. A semagram is a conceptual structure that describes a lexical entity on the basis of a wide range of characteristics, defined …

Content truncated to 500 words