The vast majority of verbs occur in more than one syntactic context. For example, the verb "know" can occur with both "that"-clauses – as in "John knows that Mary went to the store" – "whether" questions – as in "John knows whether Mary went to the store" – or noun phrases – as in "John knows Mary".
The syntactic contexts a verb can show up in correlates with its meaning (Landau & Gleitman, 1985; Fisher, Gleitman & Gleitman 1991; Lederer Gleitman & Gleitman 1995). In fact, quite fine-grained distinctions – such as those between mental state, speech, and desire verbs – are tracked by syntactic distributions (White, Dudley, Hacquard & Lidz, to appear). This fine-grainedness is interesting because it suggests that syntactic context could be useful in learning subtle distinctions that do not have direct sensory correlates, such as that between thoughts and desires.
But syntactic distribution in the abstract still only provides semantic information of a certain grain-size; syntax might tell us that a verb is a mental state, speech, or desire verb, but not much more. Recent work in both computational semantics (Sun and Korhonen 2009) and word-learning (Yuan, Fisher, Kandhadai & Fernald, 2011) suggests that this grain-size could be ground down by taking into account the selectional preferences of a verb (Resnik, 1996). For example, "eat" and "drink" are nearly identical in their syntactic distributions but differ markedly in their selectional preferences: "eat" takes direct objects that denote solids and "drink," liquids.
In this talk, I will address how incorporating selectional preferences could assist learners in discovering fine-grained aspects of mental state, speech, and desire predicate meanings. I will take as a case study "content noun" selection. Content nouns, which denote artifacts, speech acts, mental states, etc. that have propositional content – like "story" or "book" – are a prime test case because (1) selectional preference for them appears to be correlated with a classification orthogonal to the one picked out by syntactic distinctions (Moulton, 2008; Uegaki, 2012; Anand & Hacquard, ms) and (2) they appear in syntactic configurations independent of the verbs that select them (Rawlins, 2012). The former is important because the orthogonality of an information source translates into increased information given other sources – such as syntactic distribution – and the latter is important because it means that the learner could infer a noun's contenthood independently of its status as an argument of a verb, alleviating the need to bootstrap the classifications of both.