Gérard Ligozat
LIMSI-CNRS, Orsay, France
Asserting that most human activities requiring intelligence are grounded in space and time is a commonplace remark. In the context of multimodal environments, spatial and temporal information has to be represented, exchanged, and processed between different components using various modes.
In particular, spatial and temporal knowledge may have to be extracted from natural language and further processed, or general knowledge involving spatial and temporal information may have to be expressed using natural language.
This makes apparent the fact that processing spatial and temporal knowledge in natural language draws upon two main domains:
In the first domain, the main goal is upon devising formalisms for representing and reasoning about spatial and temporal information in general.
In the second domain, which is of course closely related to the first, and often considered as a subdomain of application, the main focus is upon spatial and temporal information in natural language, understanding its content and meaning, and processing it.
Both aspects interact in many applications having to do at the same time with real world data and linguistic data: story understanding, scene description, route description.
In fact, despite the obvious interest and ultimate necessity of considering both spatial and temporal aspects jointly, the state of development of the branch of AI, computational linguistics or formal philosophy dealing with time is much more advanced than that of the corresponding branches which deal with space.
Understanding temporal and spatial information in natural language---or generating natural language to express temporal or spatial meanings, implies: (1) identifying the linguistic elements conveying this information (markers), (2) elucidating its semantic (and pragmatic) content, (3) devising suitable systems of representation to process this content, (4) implementing and using those representations.
The basic linguistic markers for temporal notions in many languages are verb tenses (e.g., preterite, pluperfect in English) and temporal adverbials (yesterday, to-morrow, two weeks from now).
However, tenses also express aspectual notions, which are important for understanding the implications of a given utterance: compare He was crossing the street and He crossed the street. Only the second sentence allows the inference He reached the other side of the street.
A basic property of temporal information in natural language is its deictic nature: an event in a past tense means it happened before the time of speech. Reichenbach introduced the time of speech S, time of reference R, and time of event E to explain the difference of meaning between the basic tenses in English.
Another important component is the fact that verbs behave in various ways according to their semantic class (Aktionsart). Variants or elaborations of Vendler's classification (states, activities, accomplishments and achievements) have been in common use.
The definition and study by Prior of modal tense logics has resulted in an important body of work in formal logic, with applications in philosophy (the original
motivation of Prior's work), the semantics of natural language, and computer science. The simplest versions of tense
logics use modal operators 
, 
, 
, 
. For instance, 
means that p will be true at some
future time, and 
that p will be true at all future
times; 
and 
have the corresponding
interpretations in the past.
Hence tense logic, in analogy to natural language, uses time as an implicit parameter: a formula has to get its temporal reference from the outside.
In Artificial Intelligence, both [McD82] and
[All83] introduced reified logics for dealing with
temporal information. Reification consists in incorporating part of
the meta-language into the language: a formula in the object language
becomes a propositional term in the new language. For example, p
being true during time t might be written 
. This
allows to make distinctions about different ways of being true,
as well as quantification about propositional terms. A recent survey
of temporal reasoning in AI is [Vil94].
Typically, recent work on temporal information in natural language uses some or all of the preceding tools. A great deal of work is concerned with determining the temporal value of a given sentence. Good examples are [Web88] and [MS88].
Primary linguistic markers of space are spatial prepositions (in, on, under, below, behind, above) and verbs of motion (arrive, cross). The seminal work by [Her86] showed that prepositions cannot be analyzed in purely geometric terms, and that their interpretation depends heavily on contextual factors. Following initial work by [Van86], Vieu, Aurnague and Briffault developed general theories of spatial reference and interpretations of prepositions in French (see references in [COS93]). It appears that spatial information also involves:
Developing systems for dealing with spatial information is best understood in the larger context of spatial reasoning in AI.
A general computational framework for expressing temporal information is in terms of binary constraint networks: Temporal information is represented by a network whose nodes are temporal entities (e.g., intervals), and information about binary relations between entities is represented by labels on the arcs.
In Allen's approach, such a qualitative network will represent a finite set of intervals, and labels on the arcs will be disjunctions of the thirteen basic relations (representing incomplete knowledge). Basic computational problems will be:
The first two problems are NP-hard in the full algebra of intervals, whereas they are polynomial in the case of time points. Recent results of [NB93] identify a maximal tractable subset.
Most algorithms in this framework are variants of the constraint propagation method first introduced in this context by Allen.
Binary constraint networks also are a suitable representation for representing quantitative constraints between time points. A case in point are time maps used by Dean and McDermott (see [Vil94]).
A promising direction of research in the domain of temporal information in natural language is concerned with the integration of the textual, or discourse level. Two basic aspects are:
Typical applications include the generation of route descriptions in natural language, the maintenance and querying of spatial databases, the interpretation and generation of maps describing spatio-temporal phenomena.