Hauptseminar "Underspecification"

Alexander Koller, Stefan Thater

Block seminar: Mon 10 April to Fri 14 April 06, 9-13
Introductory session: Thursday 16 February, 13:00, room 2.11

Hauptseminar for BSc and MSc students

A key task in computational semantics is semantics construction: given a sentence, derive a semantic representation for this sentence, such as a formula of first-order or higher-order logic. An important challenge in semantics construction is that we have to deal with ambiguities. The particular class of ambiguities we are concerned with in this seminar are scope ambiguities, as in the sentence below:

(1) Every student attends a course.

This sentence can be interpreted in two ways: under the first interpretation, there is a course, say, the semantics course, which is attended by every student; under the second interpretation, every student attends some course, but not necessarily the same one.

While sentences like (1) look harmless and somewhat artificial, scope ambiguities may pose genuine processing problems: The number of readings can grow exponentially with the number of quantified noun phrases (and other scope bearing operators such as negations or modals), and large-scale grammars can predict billions of readings for real-life sentences from corpora.

Scope underspecification is the current standard approach to dealing with scope ambiguities. Its key idea is to derive from a syntactic analysis not all possible semantic readings immediately, but just one compact underspecified representation (USR). The individual readings can be recovered from the USR, but the USR could also be used as a platform for taking background knowledge and preferences into account, in order to exclude readings that were not meant in the actual context, without ever enumerating them. The field has made considerable progress over the past ten years, and efficient and stable tools are becoming available.

In this seminar, we will look at important formalisms and recent developments in scope underspecification. We will start with a brief overview over the problem of scope ambiguity and more traditional approaches to dealing with them. Then we will look at three underspecification formalisms in detail: Hole Semantics, Minimal Recursion Semantics, and Dominance Graphs. Finally, we will look at various advanced topics and techniques for dominance graphs: solving (i.e. computing the different readings), equivalence results for different underspecification formalisms, and inference techniques on USRs.