HPSG 2006 The 13th International Conference on Head-Driven Phrase Structure Grammar Tutorials and Workshop

 24-27 July, 2006, Varna, Bulgaria

Tutorial 1

Semantics within and beyond HPSG: LRS and LTAG
(Constraint-based Computational Semantics: LTAG and LRS)
Frank Richter and Laura Kallmeyer


Semantics has for the longest time been treated as a stepchild of constraint-based linguistic grammar frameworks. While it has always been acknowledged that some kind of semantic framework must ultimately be integrated into realistic grammars, constraint-based linguistics has often remained silent when it comes to making specific assumptions, or has resorted to superficial sketches of semantic structures. It has been suggested that either a ‘proper’ semantics needs to be worked out later, or the preliminary semantic representations should be thought of as an intermediate structure serving to hide the ‘real’ (and non-extant) semantics. According to this view, the current semantic representations are merely a feasible solution for practical tasks.

We regard this situation as very unfortunate and believe that constraint-based linguistics, far from being a mere recipient of ‘alien’ developments, has much to contribute to core areas of semantic theorizing. To give substance to this claim, this tutorial will discuss two theories of natural language semantics which have been worked out in considerable detail in Lexicalized Tree Adjoining Grammar (LTAG) and in HPSG, respectively. We will refer to them as LTAG semantics (Kallmeyer & Romero) and Lexical Resource Semantics (LRS, Richter & Sailer). The comparison of these two approaches to constraint-based semantics will sharpen our understanding of the range of possibilities of semantic theories in this setting and will help to distinguish their essential from their accidental properties. Among the interesting topics which arise in LTAG semantics and LRS when compared to more traditional Montague-type semantics are: (1) alternative techniques of semantic composition, (2) context effects in semantic composition, (3) new approaches to compositionality in semantics, (4) semantic underspecification, and (5) computational properties of semantic composition, in particular interleaved syntactic and semantic processing in parsing.

At first glance, one notices striking differences between LTAG semantics and LRS, which are mostly due to the differences in their underlying frameworks: LTAG is a lexicalized approach characterized by an extended domain of locality, while HPSG provides a very expressive logical description language which supports the flexible formulation of general grammar principles. Furthermore, while the LTAG approach to semantics assumes a separate level of underspecified semantic representations, LRS uses the HPSG description logic for semantic specifications. However, in spite of these differences, there are also fundamental similarities between LTAG semantics and LRS, in particular with respect to the basic concepts underlying specific analyses and with respect to the use of higher order type-theoretic languages as a means of specifying truth conditions. This is also one of the main reasons why a comparison of the two leads to a better mutual understanding and to new insight concerning the nature of these respective theories.

After an initial introduction to the two frameworks, the tutorial will focus on a comparative study of concrete analyses of particular linguistic phenomena, while demonstrating the advantages of doing semantics in a constraint-based linguistic theory. The comparison will reveal a very close relationship between the two approaches concerning the way basic assumptions about semantic mechanisms are implemented within the two frameworks. At the same time, we will identify fundamental differences in the semantic architecture due to more general formal differences between the two host frameworks, HPSG and LTAG. In this way, we will establish relations between the characteristic mathematical properties of the formalisms and the consequences for the possible semantic analyses.