Corpus-based methods and cognitive semantics:
The many senses of to run
*
Stefan Th. Gries
Abstract
The first major part of this paper is a comprehensive cognitively-oriented analysis
of the senses and their interrelations of the verb to run along the lines of much
recent cognitive work on polysemy. In the second major part, all occurrences of to
run from the ICE-GB and the Brown Corpus are coded for a variety of linguistic
parameters (so-called ID tags), yielding a complete behavioral profile of this verb.
On that basis, the paper then discusses several case studies of how such corpus-
linguistic quantitative methods can provide objective empirical evidence suggest-
ing answers to some notoriously difficult problems in cognitive linguistics; these
include the issue of prototype identification, the (degree of) sense distinctness, the
structure of the hypothesized network as well as possibilities of automatic sense
identification.
Keywords: polysemy; word sense (disambiguation); behavioral profile; semantic
network; cluster analysis.
1. Introduction
The present paper is concerned with word senses from the perspective of
cognitive linguistics on the one hand and corpus-linguistics as well as cor-
pus-based lexicography on the other hand. While many recent cognitive-
linguistic approaches to polysemy have concerned themselves with
polysemous words as network-like categories with many interrelated senses
(with varying degrees of commitment to mental representations), corpus-
linguistic approaches have remained rather agnostic as to how different
word senses are related and have rather focused on distributional character-
istics of different word senses. This paper attempts to bridge the gap be-
tween these two approaches by demonstrating how cognitive linguistics can
benefit from methodologies from corpus linguistics and computational
Stefan Th. Gries
58
linguistics; it is therefore a plea for more corpus linguistics in cognitive
linguistics and structured as follows: Section 2 provides a by necessity very
brief overview of cognitive-linguistic approaches towards polysemy and
some of their weaknesses (cf. Section 2.1) as well as some corpus-based
approaches (cf. Section 2.2). The review can of course not do justice to the
large number of studies on polysemy and especially word sense disam-
biguation; it merely serves to discuss how the problems of identifying the
different senses of a polysemous word have been addressed. Section 3 dis-
cusses the senses of the highly polysemous English verb to run on the basis
of British and American corpus data. Section 4 constitutes the central part
of this study. It introduces and exemplifies a few methodologies which
increase the descriptive adequacy of cognitively-oriented analyses of lexi-
cal items as well as resolve some notoriously difficult questions within the
cognitive paradigm. Finally, Section 5 concludes with some further exten-
sions.
2. Distinctions between senses and the relations between them:
A short review
2.1. Cognitive-linguistic approaches
One of the central areas of research within cognitive linguistics has been
the investigation of polysemy of lexemes and constructions. Traditionally,
the idea that a word is polysemous entails that the particular lexeme under
investigation (i) has more than one distinct sense (otherwise the lexeme
would be considered vague) and (ii) that the senses are related (otherwise
the lexeme would be considered homonymous).
1
The former point is usually made on the basis of a variety of well-
known ambiguity tests including the logical test, the linguistic (do so) test
and the definitional test (cf. Geeraerts [1993], Cruse [1986] and Kilgarriff
[1997] for detailed discussion). However, these tests often yield mutually
contradictory results, which is why cognitive linguists have often posited a
continuum of semantic distinctness ranging from clear cases of homonymy
on the one hand to clear cases of vagueness on the other hand; cases of
polysemy were then located somewhere between these two extremes (cf.,
e.g., Tuggy [1993] or Croft [1998]). Thus, the distinctness of different
senses of a lexeme is considered a matter of degree. Although it is probably
fair to say that cognitive linguists have focused on the analysis of how dif-
Corpus-based methods and cognitive semantics
59
ferent senses of a word are related to each other, they have of course also
been aware that the motivation of sense distinction is a non-trivial issue
since the links between senses can only be discussed once the distinctness
of senses has been established. Thus, a variety of different approaches have
been proposed to deal with this problem; let us briefly consider some ex-
amples.
Consider, as a first example, some early studies such as Brugman
(1981), Norvig and Lakoff (1987), Lakoff (1987) and Brugman and Lakoff
(1988). On the basis of intuition data, nearly every usage event minimally
different from another one constitutes a different sense. For instance,
Brugman and Lakoff argue that “a polysemous lexical item is a radial cate-
gory of senses” (1988: 478) and they posit different schemas of the English
preposition over, which often differ only with respect to properties of the
landmark. For instance, in (1a) the landmark (the hill) is vertical whereas,
in (1b), it (the yard) is not (Brugman and Lakoff’s [1988: 482–483] exam-
ples).
(1) a. The plane flew over the hill ĺ schema 1 (above and across):
vertical extended landmark, no contact
b. The bird flew over the yard ĺ schema 1 (above and across):
non-vertical extended landmark, no contact
This so-called full-specification approach (cf. Lakoff 1987) has been criti-
cized for its methodological vagueness (resulting in the high degree of
granularity – i.e., minimally different senses – pointed out above), its
vagueness of representational convention and its lack of clarity concerning
the linguistic and cognitive status of its network architecture (cf. Sandra
and Rice [1995] for discussion and exemplification), and other approaches
have been adopted to resolve this question on a principled, non-arbitrary
basis. For example, Sandra and Rice (1995) as well as Rice (1996) argue in
favor of (prepositional) polysemy on the basis of different experimental
results. As another alternative, Tyler and Evans (2001) develop a princi-
pled-polysemy approach in which a distinct sense of over is only posited iff
the meaning of over in one utterance involves a different spatial configura-
tion from over in another utterance and cannot be inferred from encyclope-
dic knowledge and contextual information.
2
However, not all these approaches are equally useful. For example, it is
unclear whether the results of the sorting tasks of Sandra and Rice (1995)
or Rice (1996) can actually be attributed solely to semantic differences of
Stefan Th. Gries
60
the uses (which also undermines the results’ utility in refuting monosemy
approaches): unlike recent experimental work by, say, Klein and Murphy
(2001, 2002), the experimental sentences were not balanced with respect to
all lexical items contributing to subjects’ decisions. Moreover, different
distance measures and clustering algorithms result in different amalgama-
tion schedules and different degrees of granularity, but Sandra and Rice do
not provide such details, which makes the evaluation of their findings diffi-
cult.
It is only very recently that cognitive linguists have turned to corpus
data as a source of evidence for sense distinctions. For example, Croft
(1998: 169) argues in favor of investigating the distinctness and conven-
tionality of senses corpus-linguistically. He points out how semantically
different direct objects of to eat correlate with uses distinct in terms of the
arguments they occur with. In addition, Fillmore and Atkins’s (2000) dis-
cussion of to crawl is cognitive-linguistic in the sense that the relations
between different senses of to crawl are motivated both experientially and
frame-semantically, but also truly corpus-based as it relies on an exhaustive
analysis of a complete concordance. Finally, Kishner and Gibbs (1996) (as
well as Gibbs and Matlock [2001]) discuss associations (of unmentioned
strengths) of different senses of the English adverb just and to make on the
one hand to different R1 collocates (i.e., words at the first slot to the right
of the word of interest) and syntactic patterns on the other hand. They dem-
onstrate “that people’s choice of a sense of just is in part determined by the
frequency of co-occurrence of particular senses of just with particular
classes of words” (1996: 27–28) as well as situational characteristics, which
results in some resemblance to a frame-semantic approach. Lastly, they
propose that such results generalize to (words of) other syntactic categories,
e.g. the verb to run and, in Gibbs and Matlock (2001: 234), argue that “if
polysemous words are best described in terms of lexical networks, then our
findings suggest the need to incorporate information about image schemas
and lexico-grammatical constructions in drawing links between different
senses of a polysemous word”, a proposal to which we will return.
2.2. Corpus-based approaches
Especially the last approach by Kishner and Gibbs bridges the gap between
cognitively oriented approaches and the linguistic paradigm in which the
question of how to determine whether two uses of a particular word instan-
Corpus-based methods and cognitive semantics
61
tiate two different senses or not has probably received most attention,
namely (corpus-based) lexicography; we will turn to this approach now.
Organizing and formulating a dictionary entry for a word requires many
decisions as to whether two citations of a word instantiate senses differing
enough that the word’s entry needs to be split or whether the citations in-
stantiate senses similar enough to be lumped
together. Although the lexi-
cographer’s interest in sense distinctions need not coincide with that of
linguists of a more theoretical persuasion, the basic question of course re-
mains the same. Given these questions, recent lexicographic work has ar-
rived at the conclusion that word senses as conceived of traditionally do not
exist and has therefore adopted an increasingly corpus-based approach. For
example, Kilgarriff (1997: 92) argues in favor of “an alternative conception
of the word sense, in which it corresponds to a cluster of citations for a
word”. In the simplest possible conception, “corpus citations fall into one
or more distinct clusters and each of these clusters, if large enough and
distinct enough from other clusters, forms a distinct word sense” (Kilgarriff
1997: 108). According to him, much lexicographic work more or less con-
forms to the following characterization: first, call up a concordance for the
word. Then, divide the concordance lines into clusters which maximize
intra-cluster similarity and minimize inter-cluster similarity. Third, for each
cluster, identify what makes the member of a cluster belong together (and
change clusters where necessary), and finally, encode these conclusions in
lexicographese (cf. also Biber [1993] and Hanks [1996: 82]). Similarly,
Hanks (2000: 208–210) argues for a focus on separate semantic compo-
nents (jointly constituting a word’s meaning potential), which can be
weighted in terms of their frequency and predictive power for regular word
uses.
However, the above is only a very abstract idealization of the actual
cognitive processes underlying sense identification and distinction. This
and the fact that many of these processes result in apparently subjective
decisions is immediately obvious once a user consults different dictionaries
on the same word (cf. Fillmore and Atkins [2000] or Gries [2001, 2003a]
for discussion). Therefore, corpus-based lexicographers have begun to for-
mulate strategies to provide a more objective foundation for resolving such
issues by, for instance, identifying corpus-based traces of meaning compo-
nents etc. In order to bring together both cognitive-linguistic and corpus-
based lexicographic approaches, it is necessary to briefly review the two
lexicographic approaches upon which the present approach relies most.