A perennial problem in literature on word-learning addresses how children learn the meanings of words like “think” and “want”. Since the mental states they refer to are unobservable—-one can’t point to thoughts or desires even if they can see their effects—-it is unclear how a learner could know to link “think” to the concept thinking and “want” to the concept wanting. One proposal for how children learn these meanings is that they can categorize words by the sorts of sentences they show up in (Gleitman, 1990). For instance, “think” can show up in sentences like (1a) but is odd in sentences like (1b) and (1c) (denoted by the “*”) while “want” can show up in sentences like (2b) and (2c) but is odd in sentences like (2a). 1a. Carla thinks that Janet went to the store. 1b. *Carla thinks Janet to go to the store. 1c. *Carla thinks a piece of cake. 2a. *Carla wants that Janet went to the store. 2b. Carla wants Janet to go to the store. 2c. Carla wants a piece of cake. Here, I present evidence for this proposal—-that features of a sentence carry information about the meaning of words in that sentence—-using both experimental methodologies and computational modeling. I show that there exists a nontrivial correlation between the sentences a word can show up in and that word’s meaning.