Grounded models of meaning
Hauptseminar for BSc and MSc students
Alexander Koller and Andrew Gargett
Wednesdays, 8:30 - 10:00; C7.1, room U.15
There is an old semanticist joke which says that the meaning of life is the constant life'. Indeed, traditional approaches to semantics, such as Montague Grammar, have focused on representing the meaning of words with symbols in a logical language. These symbols are not inherently more meaningful than the word strings themselves, and only receive an interpretation through world knowledge axioms which relate them to other such symbols. Lexical semantics can be seen as an effort to provide such axioms, either in terms of logical meaning postulates or of statistical relationships derived from corpora.
Recently, there has been quite a bit of research which takes a different route: In the presence of a concrete domain, one can directly map words or logic symbols to individuals or concepts in that domain, and say that the meaning of the words is grounded in the domain. The domain may be as concrete as the real world or a movie file, or as abstract as reports of baseball games or a collection of weather data. This makes the meaning of a word very tangible: whenever we encounter a new object in the domain, we will be able to see whether it is described by the word. It is not only unnecessary to axiomatize the relations between words or concepts explicitly; instead, such relations could potentially be learned from grounded meaning models and observations in the domain. In this seminar, we will read and discuss some of the recent literature on this topic.