Greg Detre
Tuesday, October 08, 2002
see
scribbles on printout
http://coral.lili.uni-bielefeld.de/Classes/Winter98/ComLex/GibbonElsnet/elsnetbook.dg/node15.html
Lexical semantic microstructure, as in Pustejovsky's Generative Lexicon Theory (for feature structure details, see [Pustejovsky 1995]).
� Qualia structure (semantic properties),
� Event Structure (temporal properties),
� Argument Structure (predicate-argument relations),
�
Inheritance structure (generalisation macrostructure).
http://budling.nytud.hu/~kalman/reading/siggen94/node5.html
GLT can be briefly characterized as a system which involves four levels of representation which are connected by a set of generative devices accounting for a compositional interpretation of words in context, namely: the argument structure which specifies the predicate argument structure for a word and the conditions under which the variables map to syntactic expressions; the event structure giving the particular event types such as S (state), P (process) or T (transition); the qualia structure distributed among four roles FORM (formal), CONST (constitutive), TELIC and AGENT (Agentive); and the inheritance structure which involves two different kinds of mechanisms:
� the fixed inheritance mechanism, which is basically a fixed network of the traditional isa relationship found in AI, enriched with the different roles of the qualia structure;
� the projective inheritance mechanism, which can be intuitively characterized as a way of triggering semantically related concepts which define for each role the projective conclusion space ( PCS). For instance in the PCS of the telic and agentive roles of book we will find at least the following predicates: read, reissue, annotate, ... and write, print, bind, ... (respectively).
The most important of the generative devices connecting these four levels is a semantic operation called type coercion which ``captures the semantic relatedness between syntactically distinct expressions'' (Pustejovsky, 1994a). Another notion introduced is that of lexical conceptual paradigms ( LCPs), as formalized in (Pustejovsky, 1994b). We will say that the aim of an LCP is to capture the conceptual regularities across languages in terms of cognitive invariants, like ``physical-object'', ``aperture'', ``natural kind'' and alternations such as ``container/containee'', etc. Moreover, the possible syntactic projections are associated with LCPs. For instance, one can say ``I left a leaflet in/inside the book at the page I want you to read' as book is an information-phys_obj-container whereas for instance one cannot say ``I put the book in the top of the table'' as ``the top of the table'' is a surface and not a container.
In the following, we will focus on two basic mechanisms of GLT, which allow us to bridge the word usage gap, that is, on a scale of lexical specificity, from free-combining words to idioms. These are:
(1) Reference to the qualia structure: By giving every category the ability to make reference to specific semantic functions, we are encoding the ``semantic basis'' of word usage information with a lexical item. This gives rise to semantic collocations.
(2) Cospecification: This is the basic means of encoding specific usage information in the form of either coherent argument subtypes, or already lexicalized phrases, giving rise to idiosyncrasies and idioms, respectively.
http://www.cs.brandeis.edu/~llc/genlex.html
The focus of research in Generative Lexicon Theory is on the
computational and cognitive modelling of natural language meaning. More
specifically, the investigation is in how words and their meanings combine to
make meaningful texts. This research has focused on developing a lexically
oriented theory of semantics, based on a methodology making use of formal and
computational semantics. That is, we are looking at how word meaning in natural
language might be characterized both formally and computationally, in order to
account for both the subtle use of words in different sentences, as well as the
creative use of words in novel contexts. One of the major goals in our current
research, therefore, is to study polysemy, ambiguity, and sense shifting
phenomena in different languages. .
http://www.cs.brandeis.edu/~federica/articles/book-summary.html
The framework of Generative Lexicon theory (Pustejovsky, 1995) is built on the assumption that lexical items reflect the creative aspect of language use and are able to combine to create new meanings in different contexts. � Much of the earlier research on the generative properties of the lexicon has provided substantial evidence that the study of word meaning requires a richer notion of lexical representation and compositionality, which avoid an unbounded enumeration of lexical senses for polysemous words. Taking these assumptions as the starting point, we intend to present the subject of word meaning and creative use of language from two angles. One which provides additional evidence for the generative lexicon approach, the other which discusses its repercussions and its pertinence to other domains.
http://www-users.cs.york.ac.uk/~mdeboni/research/generative_lexicon.html
Pustejovsky's lexicon model attempts to explain semantic problems such as the polymorphic nature of language and the creative use of words in novel contexts.
This is done by organising words into a generative lexicon of core word senses. This core set is then combined through a number of "generative devices" to obtain the larger set of word senses that makes up the lexicon of a language.
The idea of a generative lexical model is contrasted to a more usual sense enumerative lexicon, where each word has a literal meaning and lexical ambiguity is treated by multiple listing of words. The enumerative lexicon approach fails to explain a number of linguistic phenomena:
According to the
generative lexicon approach, lexical items are decomposed into structured forms
(or templates) that provide the framework for the composition of lexical meanings.
A generative lexicon
has at least four levels of semantic representations:
The semantics of a
lexical item a can therefore be
defined as a structure composed by the following components:
a = < A,
E, Q, I >
where A is the argument
structure, E is the specification of the event type, Q provides the binding of
A and E in the qualia structure and I is an embedding transformation that
determines what information is inheritable from the global lexical structure.
The argument structure
is made of:
The event structure
defines the order in which events occur (both sequential and non-sequential)
and the relation between an event and its subevents. This structure permits a
thorough analysis of meaning of verbs, distinguishing, for example, between
accomplishment verbs and achievement verbs.
Qualia structure is
derived in part from the Aristotelian view of word meaning which identified
"modes of explanation" (aitiai).
In particular, the following aspects of meaning (qualia) are identified:
Semantic types may be
combined to become a complex type through the use of the "dot"
operator. A "type constructor" may create a complex type for a term a
which carries senses s1 and s2, thus yielding a lexical paradigm containing the
senses {s1.s2, s1, s2}, i.e. the combination of s1 and s2, s1 and s2. Thus, for
example, a noun such as door, which is made of the combination of the types
phys_obj and aperture will have all the following types available for
expression: {phys_obj.aperture, phys_obj, aperture), thus illustrating the fact
that doors are sometimes referred to simply as physical objects, sometimes as
apertures, and sometimes as a combination of the two.
The levels of semantic
representation are connected by a set of generative devices which provide the
compositional interpretation of words in context.
These generative
devices include the following semantic transformations:
Lexical inference is
not examined in any detail by Pustejovsky, but he suggests that coercion and
co-composition could be used as a type of "enthemymic" inference, a
type of inference first analysed by Aristotle in his Rhetoric. An enthymeme is
a set of two propositions added as argument, such that the addition of a third
proposition results in a categorial syllogism, e.g. "James is a Texan,
therefore he is tall", where the additional proposition is "all
Texans are tall". This could be applied to the generative lexicon, for
example in a phrase like "Steven King began a new novel", with the
inference "Steven King began to write a new novel" given by the
information contained in the noun phrase "Steven King" and the
"agentive" value of novel which is (informally) "someone writes
a novel".
However Pustejovsky
merly suggests the possibility of inference, without examining the problem in
any depth.
The lexicon presented
by Pustejovsky appears to contain a considerable amount of information that a
linguist might consider better suited to a knowledge base, and the distinction
between linguistic knowledge and commonsense knowledge appears to be very
"fuzzy". However, it is argued that although there appears to be a
continuum between the two types of knowledge, in some cases linguistic
behaviour is best treated as language specific knowledge, not in terms of
general inference mechanisms.
One major drawback of
the generative lexicon approach is that it is not at all clear how the
information in the lexicon would be gathered. In common with most generative
linguistics approaches, Pustejovsky does not draw his examples from an analysis
of the English language "as spoken", i.e. he does not systematically
analyse the linguistic data available, but rather seems to use his
"intuition" as a native speaker to decide what a native speaker would
or would not do (see, for example, Seuren 1998, pp.260ff for a critique of the
Chomskian approach to empirical evidence). It is unclear if a systematic
approach to the writing of a generative lexicon would be possible, and even the
theory is incomplete (for example, the lexical inheritance structure is not
examined in any detail).
This approach, however,
does look promising, and certainly allows for a very rich semantic lexicon
which explains how objects may be "viewed" systematically from
different angles (e.g. a book may be something to read or a physical object
that falls on one's foot). The idea of using the lexicon in order to make
inferences on given propositions, even though only mentioned in passing,
appears very interesting and with much potential for applications such as
information extraction where it is necessary to infer as much as possible from
the given sentences.
A partial
implementation of the generative lexicon approach is given by [Buitelaar 98]
who automatically generates an ontology and semantic database of semantic types
from Wordnet [Miller et al. 1990]. Buitellar's method attempts to generate a
set of systematically related senses constituting basic "types" which
are then connected to produce more complex types. No further anaysis is made of
these basic types, however, and the characteristics of the types (argument
sturcture, event structure, qualia structure, lexical structure) are ignored.
Thus far, Pustejovsky seems to be following in Quillian�s footsteps, whose semantic networks were an attempt to represent lexical memory as �a complex network of elements and associations interconnecting them�. In drawing this parallel, we can see one of the major contrasts between Pustejovsky�s approach and Wordnet�s, in that he is actively trying to relate semantic information to syntactic form.
However, his ideas in this paper are heavily based on earlier ideas in the field of Generative Lexicon Theory, which he fails to adequately introduce.
. As the extremely helpful discussion at:
http://www-users.cs.york.ac.uk/ ~mdeboni/research/generative_lexicon.html
raises, Pustejovsky�s model blurs the distinction between a linguistic knowledge base and commonsense knowledge
join semi-lattice???
psycholinguistic evidence???
formalise/add-on within WN???
how much detailed work is necessary for the whole English language??? might the process be algorithmised after a certain point, or done statistically???
does it allow multiple inheritance??? complex/dot object???
could even this system actually capture our concepts in all their glory, or does something about language add to the concepts it expresses/is based on???
do you rely on anything more than intuition in your types etc.???
e.g. entity/event/quality, or the Aristotelian 4-part qualia structure�
what�s the difference between a type structure for concepts and for natural language???
apart from being a bit differently categorised, can we see this as essentially a derivation of Quillian�s semantic nets???
is this GL??? what (else) is GL??? what else does he say in his book???
complex vs functional???
linguistic generalisations vs metaphysical considerations???
how would he represent: colours, numbers, fucntion words, prepositions, adverbs, linguistic relations (e.g. the concept of a �sentence�)
dictionary
vs knowledge base vs common sense reasoning???