Computational Semantics and Pragmatics for Natural Language



In specific contexts, natural language utterances carry remarkably precise content. Consider the last sentence in this discourse:
I want to hold a barbecue. Some vegetarians may be coming. What can I do for them?
A general gloss for it is unsatisfying:
What actions for me to take towards vegetarians are logically possible?
It means something much more specific:
Assuming vegetarians come to my barbecue, what actions will be available to me (in virtue of the properties of barbecues and vegetarians in the actual world) that will contribute to making the event a success?
Computational semantics is devoted to the specification of an implemented theory that would describe correct and precise meanings, such as these, for utterances in context.

Computational semantics shares with formal semantics research in linguistics and philosophy an absolute commitment to formalizing the meanings of sentences and discourses exactly. The difference among these fields reflects their overall enterprises. Linguistic semantics, for example, is looking for an account of human knowledge of meaning that accounts for crosslinguistic variation and human language learnability. Philosophical semantics aims to situate knowledge of meaning within a general understanding of the intentionality of human mental states. The distinctive concerns of computational semantics include the following questions:

  • How can ambiguities and contextual dependencies in the meanings of sentences be represented compactly?
  • How can they be resolved automatically and efficiently?
  • How can semantic representations be related to other computational representations of the world?
  • How can we compute the inferential consequences of semantic representations?
These concerns lead to a focus on ontology and reference within computational semantics.
  • Our focus on ontology responds to the fundamental challenge of interpreting language in terms of the language-independent concepts of some underlying domain of discourse.
  • Our focus on reference reflects the fact that the bridge between language and the world is not accomplished directly, by predefined semantic conventions, but rather is mediated by speakers' particular intentions in particular utterances to evoke particular elements of an underlying ontology.
This focus identifies a level, suggested by Bonnie Webber and developed in joint work with Ali Knott, Aravind Joshi and myself, in which discourse is organized into multi-clause descriptions of entities, including individuals, sets, eventualities, situations, etc. This level of discourse exploits the same syntactic and semantic mechanisms as the clause; its meaning derives from familiar semantic principles: anaphoric presupposition, compositional semantics, and defeasible inference. Key questions for this perspective are then: What kinds of entities do discourses describe? And, what kinds of linguistic, discourse and inferential resources mediate the description? Obviously, there is close connection between this perspective and the perspective of my generation research. This is no accident.

Stone 92 Matthew Stone. Or and Anaphora. Proceedings of SALT 2, 1992, pages 367--385.
This paper argues that an analysis of pronouns as descriptions is required to account for sentences with disjunctive split antecedents. Here is one of the catchy examples:
It's interesting what happens if a man calls a woman or a woman calls a man. Sure, they're nervous about making the call, and they're suprised to get it. But even today, she waits for him to ask her out.
Today, I might take a more proof-theoretic view.
Stone 94 Matthew Stone. The Reference Argument of Epistemic Must. Proceedings of IWCS 1, 1994, pages 181--190.
This paper argues that an utterance of must p refers to a salient, justified argument in the context which supports p. It is the strength of the argument and the speaker's intention in referring to it that accounts for the strengh of must in some contexts:
Now [Af] and [Af] must both be tangent points on the T component in the f-plane; otherwise by Lemma 1 the component would extend beyond these points.
and the weakness of must in others:
The handsome bird was solitary; its mate must be at home, silently guarding the nest.
Today, new research on presupposition in semantics and argumentation in artificial intelligence could be used to make the case stronger and more precise.
Stone and Hardt 97 Matthew Stone and Daniel Hardt. Dynamic Discourse Referents for Tense and Modals. Proceedings of IWCS 2, 1997, pages 287--299.
Tense and modality are often thought to be anaphoric. In this paper, we argue that tense and modals, just like all other discourse anaphors, participate in strict-sloppy ambiguities under deletion:
John would give slides if he had to give the presentation. Bill would just use the chalkboard.
We apply Hardt's theory of dynamic discourse referents to give such ambiguities an account that exactly parallels Hardt's treatment of pronouns and verb phrase ellipsis.

This paper has been revised for the conference proceedings book. Here is the original version that appeared at the conference.

Stone 99 Matthew Stone. Reference to Possible Worlds. RuCCS Report 49, Rutgers University, April 1999.
In modal subordination, a modal sentence is interpreted relative to a hypothetical scenario introduced in an earlier sentence:
There may be other 1961 state committee retirements come April 18, but they will be leaving by choice of the Republican voters.
In this paper, I argue that this phenomenon reflects the fact that the interpretation of modals in an anaphoric process, precisely analogous to the anaphoric interpretation of tense. Modal morphemes introduce alternative scenarios as entities into the discourse model; their interpretation depends on evoking scenarios for described, reference and speech points, and relating them to one another.

The current version is a revision of an earlier paper, The Anaphoric Parallel between Modality and Tense. IRCS Report 97-06, University of Pennsylvania, May 1997. That paper is also available directly from Penn

Webber et al. to appear Bonnie Webber, Matthew Stone, Aravind Joshi, and Alistair Knott. Anaphora and Discourse Semantics.
To appear in Computational Linguistics.
We argue that many discourse relations can be explained non-structurally in terms of the grounding of anaphoric presuppositions. This simplifies discourse structure, allowing it a straightforward compositional semantics that still realizes a full range of discourse relations. This example, with its many cue words:
On the one hand, John loves Barolo. So he ordered three cases of the '97. On the other hand, because he's broke, he then had to cancel the order.
suggests the advantage of using multiple devices like structure and presupposition to account for discourse relations.
An earlier version was Discourse Relations: A Structural and Presuppositional Account Using Lexicalized TAG. Proceedings of ACL, 1999, pages 41-48.

Stone 00 Matthew Stone. Towards a Computational Account of Knowledge, Action and Inference in Instructions. To appear in Journal of Language and Computation, 2000.
I consider abstract instructions, which provide indirect descriptions of actions in cases where the speaker has key information that a hearer can use to identify the right action to perform, but the speaker alone cannot identify that action. The communicative effects of such instructions, that the hearer should know what to do, are in effect implicatures. Here's a typical abstract instruction:
Enter your name on box one of the form.
This paper sketches a computational framework for constructing and recognizing communicative intent for abstract instructions, by reasoning from a computational semantic theory.

Events I gave an invited presentation at ICOS-1: Inference in Computational Semantics to be held in conjunction with ESSLLI in Utrecht in August 1999.
Fifth International Workshop on Computational Semantics. Tilburg, NL, Jan 2003.
Organization SIGSEM the ACL special interest group on computational semantics.

August 1, 2002