The distributional hypothesis∗

Distributional approaches to meaning acquisition utilize distributional properties of linguistic entities as the building blocks of semantics. In doing so, they rely fundamentally on a set of assumptions about the nature of language and meaning referred to as the distributional hypothesis. This hypothesis is often stated in terms like “words which are similar in meaning occur in similar contexts” (Rubenstein & Goodenough, 1965); “words with similar meanings will occur with similar neighbors if enough text material is available” (Schütze & Pedersen, 1995); “a representation that captures much of how words are used in natural context will capture much of what we mean by meaning” (Landauer & Dumais, 1997); and “words that occur in the same contexts tend to have similar meanings” (Pantel, 2005), just to quote a few representative examples. The general idea behind the distributional hypothesis seems clear enough: there is a correlation between distributional similarity and meaning similarity, which allows us to utilize the former in order to estimate the latter. However, one can pose two very basic questions concerning the distributional hypothesis. The first is what kind of distributional properties we should look for, and what — if any — the differences are between different kinds of distributional properties. Looking at algorithms for distributional meaning acquisition we can discern two distinct approaches. The first is to build distributional profiles for words based on which other words surround them, as exemplified by Schütze (1992) and the Hyperspace Analogue to Language (HAL) model (Lund, Burgess, & Atchley, 1995). The second is to build distributional profiles based on in which text regions words occur, as exemplified by the Latent Semantic Analysis (LSA) model (Landauer & Dumais, 1997). These approaches are often treated as functionally equivalent when it comes to representing meaning similarities, despite the fact that they are based on different types of distributional raw materials. The second question is in what sense it is meaning that is conveyed by distributional patterns. Proponents of distributional methods often seem comfortable to ascribe meaning to distributional representations without explaining