You and me... in a vector space: modelling individual speakers with distributional semantics

Proceedings of the Fifth Joint Conference on Lexical and Computational Semantics (*SEM 2016), pages 179–188, Berlin, Germany, August 11-12, 2016. You and me... in a vector space: modelling individual speakers with distributional semantics Aur´elie Herbelot Centre for Mind/Brain Sciences University of Trento aurelie.herbelot@unitn.it Behrang QasemiZadeh DFG Collaborative Research Centre 991 Heinrich-Heine-Universit¨at D¨usseldorf zadeh@phil.hhu.de Abstract The linguistic experiences of a person are an important part of their individuality. In this paper, we show that people can be modelled as vectors in a semantic space, using their personal interaction with spe- cific language data. We also demonstrate that these vectors can be taken as repre- sentative of ‘the kind of person’ they are. We build over 4000 speaker-dependent subcorpora using logs of Wikipedia ed- its, which are then used to build distri- butional vectors that represent individual speakers. We show that such ‘person vec- tors’ are informative to others, and they influence basic patterns of communication like the choice of one’s interlocutor in conversation. Tested on an information- seeking scenario, where natural language questions must be answered by addressing the most relevant individuals in a commu- nity, our system outperforms a standard in- formation retrieval algorithm by a consid- erable margin. 1 Introduction Distributional Semantics (DS) (Turney and Pan- tel, 2010; Clark, 2012; Erk, 2012) is an approach to computational semantics which has historical roots in the philosophical work of Wittgenstein, and in particular in the claim that ‘meaning is use’, i.e. words acquire a semantics which is a function of the contexts in which they are used (Wittgenstein, 1953). The technique has been used in psycholinguistics to model various phenom- ena, from priming to similarity judgements (Lund and Burgess, 1996), and even aspects of lan- guage acquisition (Landauer and Dumais, 1997; Kwiatkowski et al., 2012). The general idea is that an individual speaker develops the verbal side of his or her conceptual apparatus from the linguistic experiences he or she is exposed to, together with the perceptual situations surrounding those expe- riences. One natural consequence of the distributional claim is that meaning is both speaker-dependent and community-bound. On the one hand, depend- ing on who they are, speakers will be exposed to different linguistic and perceptual experiences, and by extension develop separate vocabularies and conceptual representations. For instance, a chef and a fisherman may have different represen- tations of the word fish (Wierzbicka, 1984). On the other hand, the vocabularies and conceptual rep- resentations of individual people should be close enough that they can successfully communicate: this is ensured by the fact that many linguistic ut- terances are shared amongst a community. There is a counterpart to the claim that ‘language is speaker-dependent’: speakers are language-dependent. That is, the type of person someone is can be correlated with their linguis- tic experience. For instance, the fact that fish and boil are often seen in the linguistic environment of an individual may indicate that this individual has much to do with cooking (contrast with high co-occurrences of fish and net). In some …

Content truncated to 500 words