Distribution and Inference: What Philosophical and Computational Semantics can Learn from Each Other

0463407 - FLÚ 2017 RIV SK eng J - Článek v odborném periodiku
Ocelák, Radek
Distribution and Inference: What Philosophical and Computational Semantics can Learn from Each Other.
Organon F. Roč. 23, č. 3 (2016), s. 299-323. ISSN 1335-0668
Grant CEP: GA ČR(CZ) GA13-21076S
Institucionální podpora: RVO:67985955
Klíčová slova: lexical semantics * distribution * compositionality * inferentialism
Kód oboru RIV: AA - Filosofie a náboženství

Distribution of a word across contexts has proved to be a very useful approximation of the word's meaning. This paper reflects on the recent attempts to enhance distributional (or vector space) semantics of words with meaning composition, in particular with Fregean compositionality. I discuss the nature and performance of distributional semantic representations and argue against the thesis that semantics is in some sense identical with distribution (which seems to be a strong assumption of the compositional efforts). I propose instead that distribution is merely a reflection of semantics, and a substantially imperfect one. That raises some doubts regarding the very idea of obtaining semantic representations for larger wholes (phrases, sentences) by combining the distributional representations of particular items. In any case, I reject the generally unquestioned assumption that formal semantics provides a good theory of semantic composition, which it would be desirable to combine with distributional semantics (as a theory that is highly successful on the lexical field). I suggest that a positive alternative to the strong reading of the distributional hypothesis can be seen in the philosophy of inferentialism with respect to language meaning. I argue that the spirit of inferentialism is reasonably compatible with the current practice of distributional semantics, and I discuss the motivations for as well as the obstacles in the way of implementing the philosophical position in a computational framework.
Trvalý link: http://hdl.handle.net/11104/0262663