A Quantum-Like Approach to Word Sense Disambiguation

Proceedings of Recent Advances in Natural Language Processing, pages 1176–1185, Varna, Bulgaria, Sep 2–4, 2019. https://doi.org/10.26615/978-954-452-056-4_135 1176 A Quantum-Like Approach to Word Sense Disambiguation Fabio Tamburini FICLIT - University of Bologna, Italy fabio.tamburini@unibo.it Abstract This paper presents a novel algorithm for Word Sense Disambiguation (WSD) based on Quantum Probability Theory. The Quantum WSD algorithm requires con- cepts representations as vectors in the complex domain and thus we have de- veloped a technique for computing com- plex word and sentence embeddings based on the Paragraph Vectors algorithm. De- spite the proposed method is quite simple and that it does not require long training phases, when it is evaluated on a standard- ized benchmark for this task it exhibits state-of-the-art (SOTA) performances. 1 Introduction The introduction of the prototype theory by E. Rosch (1973), one of the most influential theo- ries describing concept organisation at cognitive level, completely changed the perspective in se- mantics and nowadays most of the studies in com- putational semantics consider concepts member- ship and concepts similarity as graded features in a “semantic space”. In Natural Language Processing (NLP) all the recent and fundamental studies on word and sen- tence embeddings, e.g. (Mikolov et al., 2013; Pen- nington et al., 2014; Bojanowski et al., 2016; Le and Mikolov, 2014; Sutskever et al., 2014; Kiros et al., 2015; Subramanian et al., 2018; Cer et al., 2018), as well as the older works on word spaces based on co-occurence measures, e.g. see the re- views from (Turney and Pantel, 2010; Baroni and Lenci, 2010), rely on an high-dimensional vec- tor space to represent concepts, through the appli- cation of the Harrisian distributional hypothesis, collecting contextual information from text cor- pora, and measuring their relationships by means of some kind of geometric distance. The fundamental experiments of Tversky (1977) in cognitive psychology on concept sim- ilarity and concept combination challenged the common view to adopt the Classical, Kolmoro- govian, Probability Theory (CPT) to explain and model these phenomena. Starting from Tversky’s data, a large set of newer experimental studies in cognitive psychology showed a systematic viola- tion of the basic axioms of CPT. For example, the classical problem known as the “pet-fish problem” or “guppy-effect” (Osherson and Smith, 1981) is a typical case of overextension in concept conjunc- tion: if we denote as pet-fish the conjunction of the concepts pet and fish and ask people to rate “guppy” as member of pet, fish and pet-fish, they tend to consider it as a highly typical pet-fish but neither a particularly typical pet nor fish. This vio- lates the CPT axiom stating that the probability of the events conjunction must be less or equal to the probability of the single events. In a similar way scholars in cognitive psychology proposed vari- ous experiments showing a systematic violations of CPT axioms in concept disjunction, conceptual negation, decision making and some other relevant cognitive processes: see, for example, (Hampton, 1988; Alxatib and Pelletier, 2011; Aerts et al., 2015). These studies show clearly the impossi- bility …

Content truncated to 500 words