Learning Containment Metaphors

We present a computational approach that traces the developmental process, from containment image schemas to metaphor, in four phases: a) perceptual discovery of image schemas, b) associating perceptual arguments and the relation with linguistic units, c) discovering a linguistic structure encoding the schema, and finally d) enriching the semantics of the schema via extended language usage (via a corpus). In the first three phases, we use no prior knowledge about either the perceptual or language domains; in the corpus analysis, we use the WordNet ontology. Our input is an animation based on the Heider-Simmel video, together with a small corpus of transcribed commentaries. From the image sequence, we cluster the visual angle subtended by a landmark, and find that one cluster reflects containment. This is then correlated with the sentences from the adult commentaries uttered contemporaneously with containment situations, yielding strong objectnouns and relation-preposition associations. For discovering linguistic constructs, we use no knowledge of grammatical category or syntax but find recurring patterns using the approach of (Solan, Ruppin, Horn, & Edelman, 2002). Knowing the units involved, we can identify several phrasal patterns (e.g. “X moved into”, “in the Y”). We then search a corpus with the “in the Y” schema to identify container words. We find that the most common class involving containment is location (66%), followed by group membership (20%), time, and cognition (17% each). These may be thought of as language-based non-spatial enrichments for the image schema.