ABSTRACT
Distributional approaches (DAs) to modelling meaning in use are getting more and more attention. Inspired by the distributional hypothesis of Harris (1954), they represent the meaning of a word using a vector consisting of the relative occurrence of this word with respect to other words occurring in its vicinity. Representing the meaning of a word using a vector facilitates the assessment of the similarity of meanings between words. If the distance between two word-vectors is small, then these words have similar meanings, which is to say that they occur in a given corpus with similar surrounding words at similar frequencies. One advantage of DAs is that they take some types of contextual clues into account while simultaneously lending themselves to computational implementations (Landauer et al. 1998; Lenci 2008). Distributional models of meaning were first developed to deal with word-vectors (Lenci 2018), but some expanded their scope so that they could represent a sentence-vector by combining word-vectors (Clark 2015). However, such measures of word similarity do not transfer easily to notions of sentence similarity, especially when we consider the various kinds of implicatures: compositional DAs completely omit the treatment of the implicated meaning. Nowadays, sentence similarity measures have many applications like paraphrase detection and short answer tasks (Koleva et al. 2014), automatic summarization (Erkan & Radev 2004), and machine translation (Liu & Zong 2004) and it is important to make sure sentence similarity measures take into account the full meaning conveyed by a sentence. This, as I claim in this chapter, will only be possible if both the explicit and the implicated meaning of a sentence are considered.
