Sense Abstractness, Semantic Activation and

Word Sense Disambiguation:

Implications from Word Association Norms

Oi Yee Kwong

Language Information Sciences Research Centre

City University of Hong Kong

Tat Chee Avenue, Kowloon, Hong Kong

Abstract. Automatic word sense disambiguation (WSD) often draws on a

variety of contextual cues, and decides on the most suitable sense by means of

some lexical resources or statistical classifiers. While the importance of using

multiple types of lexical information is recognised in most systems, not much

has been reported on their individual and combined effectiveness in relation to

the intrinsic nature of individual words. We attempt to address this cognitive

aspect of WSD by examining the psychological evidence regarding the internal

lexicon and its compatibility with the information available from computational

lexicons. In this study, we compare the responses from a word association task

with the lexical associations available from WordNet, to explore the effect of

sense abstractness on semantic activation, and thus the implications on the

lexical sensitivity of WSD. Preliminary results suggest that concrete senses and

syntagmatic associations are more readily activated than abstract senses and

paradigmatic associations. The results are expected to inform the construction

of lexico-semantic resources and WSD strategies.

1 Introduction

Words might have multiple meanings, resulting in word sense ambiguity. Getting the

right meaning of words in different contexts, otherwise known as word sense

disambiguation (WSD), is thus an important step in natural language processing

(NLP). Automatic WSD often draws on a variety of contextual cues, which are then

evaluated against some lexical resources or subject to statistical classifiers, to decide

on the most appropriate or most probable sense accordingly.

As Resnik and Yarowsky (1997) remarked, “disambiguation seems highly lexically

sensitive, in effect requiring specialised disambiguators for each polysemous word”.

Similarly, Ide and Veronis (1998) suggested: “… to date there has been little

systematic study of the contribution of different information types for different types

of target words. It is likely that this is a next necessary step in WSD work.” In recent

years, many research teams all over the world have gained rich experience from the

SENSEVAL workshops with their WSD shared tasks. As pointed out by Mihalcea et

al. (2004), among the 47 participating systems in the SENSEVAL-3 English lexical

sample task, “several of the top performance systems are based on combination of

Yee Kwong O. (2007).

Sense Abstractness, Semantic Activation and Word Sense Disambiguation: Implications from Word Association Norms.

In Proceedings of the 4th International Workshop on Natural Language Processing and Cognitive Science, pages 169-178

DOI: 10.5220/0002419501690178

 SciTePress

multiple classifiers, which shows once again that voting scheme that combine several

learning algorithms outperform the accuracy of individual classifiers”. However,

although the once notorious “knowledge acquisition bottleneck” is partially soothed

by statistical methods, the advancement in WSD is rarely accompanied by any

extensive account on the cognitive aspects of the lexical sensitivity of the task. Hence

the suggestion by Ide and Veronis in what might be considered a somewhat dated

source now is nevertheless still valid in a certain sense.

To better understand the contribution of different information types for different

types of target words, it is thus important to look at WSD in relation to the very

intrinsic nature of the individual words to be disambiguated (or target words), in

addition to an optimal combination of classifiers alone. We use the concept

Information Susceptibility (Kwong, 2005) to refer to the relationship between the

intrinsic features of a target word and its senses, and the effectiveness of various

lexical information to characterise them. While the intrinsic nature of a word and its

senses could comprise many factors such as frequency, abstractness, sense relatedness

and parts-of-speech (POS), in the current study we focus on the abstractness /

concreteness of individual senses, and analyse the way it corresponds to the responses

elicited from word association tests. Since word association norms are generally

assumed to reveal the organisation of our mental lexicon, they serve as a bridge

between the internal mechanism and the external modelling.

We will start with a discussion in Section 2 on the cognitive aspects of WSD from

three perspectives: introspection, psychological evidence, and computational

modelling; and how they interact. In Section 3, we present a preliminary study to

explore the effect of sense abstractness on semantic activation. In particular, we

compare the responses from a word association task with the lexical associations

available from WordNet, a widely used computational lexicon. Results and their

implications on the lexical sensitivity of WSD will be discussed in Section 4, with

future directions, followed by a conclusion in Section 5.

2 Cognitive Aspects of WSD

In this section, we will discuss the need for multiple sources of knowledge for WSD

and the evidence of the lexical sensitivity of the task from various perspectives.

2.1 An Introspective Account

Being a common and psychologically valid phenomenon of natural languages, word

sense ambiguity (or polysemy) penetrates our daily language use. Despite the apparent

non-discreteness of “sense” as Kilgarriff (1992) argued, human beings used to rely on

the predetermined senses in existing lexical resources, especially dictionaries, as a

tool for construing senses. We do not seem to have much difficulty, under normal

circumstances, to access the intended interpretation of a polysemous word in a given

context. For instance, if one says “we will have three courses for dinner”, it will be

unlikely for any hearer, not even vegetarians, to mistake it as eating up golf courses or

the grass on them.

170

Introspection alone might suggest that when processing the above sentence, we

could have been following a path like “dinner is a kind of meal, and interpreting

course as a part of a meal amongst its other possible meanings might be most

appropriate”. On the other hand, to get the meaning of courses in “he enrolled in three

courses”, instead of using paradigmatic relations such as the IS-A and PART-OF

relations employed in the previous example, we might rely more on the syntagmatic

relations between “enroll” and “course” for disambiguation.

More generally, human beings appear to use a range of cognitive strategies to

make sense of concepts of different abstractness. For instance, one can understand a

“mirror” as the instrument which reflects the image of oneself; but the simple

hypernymy relation might not be as useful for understanding an “injury”. It does not

refer to some concrete object, although it could often be visualised (such as a bleeding

wound or a broken leg). The more abstract a word, the less obvious is its external

reference, thus one could imagine how difficult it is to describe what “loss” is.

A corollary from this phenomenon is that the different strategies in making sense

of words, including the type and strength of various kinds of semantic association,

should be realised in NLP systems, especially in knowledge demanding subtasks like

WSD. Hence Quillian’s (1968) network model of semantic memory, in which the

association amongst concepts can be of very different nature, has inspired and

influenced not only the approaches in WSD but also the many computational lexicons

created for use in the task, as further discussed below.

2.2 Psychological Evidence

Quillian (1968) proposed a computational model of human memory for storing the

“meanings” of words, which remains influential in our conception of the internal

lexicon as well as in the construction of computational semantic lexicons. Apparently,

our memory stores not isolated but connected information. So a model of semantic

memory should have, in Quillian’s words, “the ability to use information input in one

frame of reference to answer questions in another”. Thus his model has a network

structure with a mass of nodes interconnected by associative links of different kinds.

The model allows two word concepts to be compared and contrasted via the links

between them. The association between concepts was found by a method generally

known as “spreading activation with marker passing”, which also underlies many

later WSD programs.

The psychological validity of such a network model is evident from subsequent

studies on lexical priming and lexical access. Priming studies (e.g. Collins and Loftus,

1975), suggest that the processing of a concept (in terms of the response time for

lexical decision tasks) would be faster if primed by a semantically related concept.

Lexical access studies (e.g. Swinney, 1979), on the other hand, propose several

hypotheses regarding the processing of multiple meanings in the case of lexical

ambiguity. There are no unanimous results, but it seems that multiple meanings are

activated at least briefly, and the influence of prior context might interact with the

nature of the individual senses, such as dominance or familiarity in terms of

frequency.

As far as the nature of the target words is concerned, the relatedness among

multiple senses could be another factor. For instance, in the lexical access literature,

171

there is a general finding that visual lexical decisions are faster for words that are

semantically ambiguous. Rodd et al. (2002) challenged the conventional assumption

that this phenomenon of “ambiguity advantage” is a result of ambiguity between

multiple, unrelated meanings, instead of multiple related word senses, with several

sets of test words more rigorously controlled for the sense relatedness therein.

Thus from the psychological perspective, we see that the access of multiple

meanings and the processing of lexical ambiguity are likely to be influenced by the

nature of individual target words. It is worthwhile to investigate how such lexical

sensitivity could be modelled in automatic WSD, and whether such modelling could

substantially benefit the latter.

2.3 Computational Modelling

Automatic resolution of word sense ambiguities has primarily depended on contextual

features, which are evaluated against some lexical knowledge sources, or subject to

statistical classifiers based on various machine learning algorithms.

In early studies, lexico-semantic knowledge for WSD was often hand-coded for

particular systems, e.g. semantic networks (Hirst, 1987), and core and dynamic

lexicons (McRoy, 1992). These are serious and rich semantic resources, but at the

expense of time, labour and scalability. With the availability of machine-readable

dictionaries, thesauri, and large corpora, researchers have explored various ways to

(semi-) automatically acquire semantic information from them.

WordNet (Fellbaum, 1998) is probably the first broad coverage general

computational lexical database. It defines word senses via synonymy, linked by

relational pointers (e.g. hypernym, antonym, etc.), forming semantic nets. It is,

however, known for the lack of syntagmatic relations, and researchers started to

address this gap with various means to enrich the lexicon with topic associations and

other broader semantic relations to enhance word access (e.g. Ferret and Zock, 2006).

Following the upsurge of corpus-based and empirical methods, statistical

approaches become the common practice in automatic WSD. Multiple knowledge

sources are modelled computationally as a variety of features from topical and local

contexts. The prevalence of machine learning approaches in WSD is evident from the

recent SENSEVAL workshops (Mihalcea et al., 2004).

Thus knowledge-based methods for WSD address the need for multiple types of

lexical knowledge by using semantic networks containing different kinds of semantic

relations (e.g. IS-A, PART-OF, thematic relatedness, etc.), and statistical methods

address the issue by getting an optimal combination of the various knowledge sources

for individual target words (e.g. Mihalcea, 2002). However, it is interesting to note

that there is somehow no comprehensive qualitative and objective account of the

relation between the disambiguation results and the nature of individual target words

underlying the apparent lexical sensitivity of the task.

2.4 When They Meet

To say that different information types contribute variably to different target words is

essentially presupposing that different types of lexical information vary in their

172

effectiveness to characterise a sense and distinguish it from other senses of the same

word. Thus it is not enough to just conceptualise senses by a certain dimension (e.g. a

certain semantic relation) across the board. The very intrinsic nature of a given

word/sense and its relation with different semantic dimensions must also be

thoroughly examined.

Leacock et al. (1998), for example, observed that “the benefits of adding topical to

local context alone depend on syntactic category as well as on the characteristics of

the individual word”. Such “characteristics” are equivalent to the intrinsic properties

of the target words in our discussion, which might include abstractness, frequency,

sense relatedness, POS, amongst others, and as we propose, are critical for

understanding the lexical sensitivity of the WSD task.

Information Susceptibility (Kwong, 2005) thus refers to the relation between the

intrinsic properties of a word and the effectiveness of various types of lexico-semantic

knowledge to characterise it. Such information is absent from existing lexical

resources. Based on the performance of a spectrum of semantic relations to

disambiguate a set of target words, it was observed that senses involving more

abstract thinking tend to be disambiguated only with broader semantic relations. This

observation also coincides with findings from human word association tests. For

instance, in the Birkbeck word association norms (Moss and Older, 1996), “loss”

triggers associations like “death” and “grief”, which cannot be related via a simple IS-

A relation, in contrast to responses like “magic” triggered by “trick” which are simply

synonymous. Hence, from the cognitive perspective, the knowledge on the

information susceptibility of individual target words is important for fine-tuning WSD

systems and informing the optimal combination of disambiguation cues. To provide

this knowledge in existing lexical resources, we need to examine the nature of target

words (in terms of frequency, abstractness, sense relatedness, POS, etc.) in the context

of lexical access and WSD.

Hence, in the current study, we focus on one aspect of the intrinsic nature of words,

namely sense abstractness, and explore how it varies with the kind of lexical

association in our mental lexicon and how it might affect the effectiveness of various

kinds of semantic knowledge in disambiguation. The study is based on data from

word association norms, and we compare the responses gathered in Hirsh and Tree’s

(2001) study with the lexical association available from the widely used semantic

lexicon WordNet. Since word association is a commonly used method to probe the

organisation and structure of the internal lexicon, and computational lexicons and

ontologies are assumed to model human conceptualisation, the comparison is

expected to allow us to better understand the human semantic repertoire and thus the

computational information demand for individual words in the lexically sensitive

disambiguation process.

3 A Preliminary Study

In this section, we present our preliminary study on the effect of sense abstractness on

semantic activation, by comparing the responses from a word association task with

the lexical associations available from WordNet. The word association responses are

assumed to be reflective of the organisation of the internal lexicon, and WordNet

173

information is used for operationalising the type and strength of semantic association

links. The objectives are two-fold: (1) to investigate the effect of one aspect of the

intrinsic nature of words on semantic activation, and (2) to study the implications of

such target-dependent semantic association patterns, if any, on the lexical sensitivity

of WSD.

3.1 Materials

We used the 90 stimulus words in Hirsh and Tree’s (2001) word association test as

our target words in this study, and focused on the top five responses elicited from the

young cohort. (Hirsh and Tree analysed the difference between the responses from

young adults and those from older adults.)

As mentioned, WordNet organises word senses in the form of synsets (i.e. sets of

synonyms) with relational pointers linking among different synsets to form some sort

of a semantic hierarchy. The synsets are also organised under 45 lexicographer files

based on syntactic category and logical groupings. WordNet was created for

psycholinguistic studies of the mental lexicon to start with but turned out to be an

electronic resource widely used by computational linguists. Thus, in this study, we

used WordNet 2.1: (1) as a dictionary to provide information on the number of senses

for a word, (2) as a computational model of the internal lexicon in the form of a

semantic network, despite its known bias toward paradigmatic relationship in general,

and (3) as a means to distinguish between concrete and abstract concepts.

3.2 Method

The 90 target words were first checked against WordNet 2.1 for the number of senses

they have, and each sense against the lexicographer files to which they belong, to

determine whether they correspond to concrete or abstract concepts.

The top five responses from the young cohort were taken and compared to two

groups of word associations obtained from WordNet. The first, which we will call

WNAsso1 below, consists of all words in the synsets (words composing the synsets

only, excluding glosses and examples) directly related to the synset(s) to which the

target word belongs. These directly related synsets include antonyms, hypernyms,

hyponyms, holonyms, meronyms, and coordinate (or sister) terms. The second, which

we will call WNAsso2 below, consists of the words in the glosses and examples in

these related synsets. Thus the first could be taken as the cluster of words

corresponding to mostly paradigmatic relations with the target words, and the second

more likely to be broader semantic relations and associations, including some

syntagmatic relations. In the current study, we only looked at the noun hierarchy, and

ignored senses of the target words under other POS.

It is, however, possible that Hirsh and Tree’s respondents might not have always responded

to a stimulus word as a noun, given that they were not specifically instructed to do so.

174

4 Results and Discussion

Among the 90 target words, 16 were monosemous and 73 were polysemous according

to the WordNet noun database.

The polysemous words have 2 to 15 senses, with an

average of 4.7 senses. There was one target word (i.e. “sly”) which was found only in

the adjective database in WordNet, and was ignored in subsequent analysis.

There are 45 lexicographer files under which the synsets are organised based on

syntactic category and logical groupings, 26 of which are relevant to noun senses. We

identified 7 concrete classes and 19 abstract classes. The concrete classes include

animal, artifact, body, food, object, person, and plant. The rest are the abstract classes,

including attribute, cognition, feeling, motive, and so on. This dichotomous

distinction may have certain limitation, which will be further discussed below.

Thus for the remaining 89 target words, all (16) monosemous words and 20 (out of

73) polysemous words only have concrete or tangible senses, 3 polysemous words

only have abstract senses (they are “bunch”, “traffic” and “wedding”), and 50

polysemous words have both tangible and abstract senses. This results in altogether

222 tangible senses and 136 abstract senses. The fact that more tangible senses are

observed is expected because Hirsh and Tree (2001) had indicated in their study that

their stimuli were “mostly names of concrete or picturable objects or likely to elicit

the name of a concrete object”. However, they did not mention with respect to which

sense the “picturability” was determined in the event of polysemy.

Table 1 shows the results for comparing the association responses with WNAsso1

and WNAsso2. The figures show the number of target words found under the various

overlapping scenarios. The overlapping could correspond to one or more senses of a

given target word. Thus WNAsso1 was assumed to contain mostly paradigmatically

related words and WNAsso2 broader associations including some syntagmatically

related words. It can be seen that for all sense types, the “WNAsso2” and “Both”

columns make up the majority, and only three cases overlap with purely paradigmatic

responses.

This is in consensus with Hirsh and Tree’s analysis, where they observed

more syntagmatic responses. There are, however, a few exceptional cases which have

none of their responses overlapping with any of our WordNet data. Some preliminary

qualitative analysis of the results is discussed below, regarding the relationship

between the numerical figures and the abstract/concrete nature of the words.

Table 1. Results for comparing the association responses with WordNet data.

Overlapping

Word Type Sense

Abstractness

WNAsso1 Only WNAsso2 Only Both None

Monosemous All Tangible 1 6 8 1

All Tangible 2 5 11 2

All Abstract 0 1 1 1

Polysemous

Both T & A 0 22 27 1

Hirsh and Tree (2001) claimed to have 41 unambiguous nouns. This was more than what we found with

WordNet senses, which might have more fine-grained senses.

Note that this observation does not preclude any syntagmatic responses for the three cases in the word

association test, which might not be found in our limited syntagmatic associations obtained from

WordNet.

175

The lexical access literature suggests that multiple senses might be at least briefly

activated in the case of polysemy, but has not systematically explored the sense

abstractness factor. For 21 out of the 50 polysemous words with both tangible and

abstract senses, the association responses overlap with WNAsso1 or WNAsso2

corresponding to one or more of their tangible senses only. For instance, the stimulus

word “zip” has four noun senses in WordNet, including “zero”, “postcode”, and

“vigour” which are abstract, and “zipper” which is tangible. The top five responses

are “trouser(s)”, “fly”, “button(s)”, “jacket” and “clothes”. All except “fly” were

found among the WordNet associations corresponding to the “zipper” sense. This is in

contrast to 6 (out of 50) with responses overlapping with WordNet associations

corresponding to their abstract senses only. For example, the stimulus word “safety”

has two tangible senses and four abstract senses in WordNet. Only the response

“security” overlaps with WNAsso1 and WNAsso2 for one of the abstract senses

referring to “a state of being certain that adverse effects will not be caused”. This

observation suggests that in the case of polysemy with both tangible and abstract

senses, the tangible concepts seem to be relatively more accessible from the internal

lexicon, assuming word association responses reflect the closest and strongest

associations in the internal lexicon.

Notwithstanding the above observation, the preference for tangible senses might

also be a result of frequency or familiarity. However, the frequency effect is not

obvious from the current study. While WordNet senses are ordered by frequency,

there is no significant pattern to show that the responses are necessarily related to the

first few senses. There are several cases where the overlapping corresponds to the top

one or two senses of a word, but no conclusive remarks could be made at this stage,

and further investigation with better control on the sense frequency would be

required.

As mentioned earlier and evident from Table 1, syntagmatic associations appear to

be more prevalent than paradigmatic ones. This is not surprising given the much

broader possibilities with syntagmatic associations. Nevertheless, about 38% of all

target words have responses overlapping with WNAsso2 only. So what underlies the

absolute dominance of syntagmatic associations in these cases? Could it be related to

the specificity and concreteness of the senses? However, looking at the six

monosemous words under this category, they are nevertheless located at a position in

the WordNet hierarchy as deeply branched as the other monosemous target words,

and thus they appear similarly specific. At the same time, the apparent inferiority of

paradigmatic responses might be an artifact of the WordNet classification itself. For

instance, the hypernym of “ankle” is “gliding joint”, and that for “kennel” is

“outbuilding”, which might be too specialised for daily usages and conception.

Hence, even though the top response for “ankle” is “foot”, they are not

straightforwardly related in the WordNet database. The concreteness hypothesis is not

supported either, given that all the monosemous target words are tangible concepts,

there is still a substantial portion of them dominated by syntagmatic responses. One

limitation, however, is that our dichotomous distinction between concreteness and

abstractness might be too coarse, whereas abstractness / concreteness could be a

continuum. Another drawback of using the lexicographer files for the distinction is

that even seemingly tangible classes like “animal” could also be abstract with words

like “Animalia”. We definitely need to address this issue in future studies.

176

Thus it seems that although paradigmatic relations like IS-A or PART-OF are an

important part of our semantic knowledge, spreading activation seems to favour

broader associations. What makes the syntagmatically related words stronger links

would be our focus in future study. In particular, we plan to extend our comparison

with corpus-based data sources, for more syntagmatic relations and general

associations. We will also refine our definition of sense abstractness. In addition to

the concrete/abstract distinction, other factors like frequency, relatedness among

senses, POS and possibly others might all contribute to the intrinsic nature of words,

and target words need better control on these dimensions in future work. Moreover,

given our preliminary findings on sense abstractness and semantic activation, one

important future direction is to further examine disambiguation performance on

concrete and abstract senses and to investigate their respective information demand

for WSD.

Thus our current preliminary study has at least the following implications on the

lexical sensitivity of WSD and the classification of senses in computational lexicon

for WSD: (1) Tangible concepts seem to be more easily activated in the internal

lexicon, and even in the case of polysemy, tangible senses appear to be more

accessible than abstract senses, although frequency and familiarity might also play a

role. (2) While paradigmatic associations form an important part of our semantic

knowledge, the observed dominance of syntagmatic associations might inform the

computational modelling of the internal lexicon, such that different weights might be

attached to different kinds of associations for words with different nature. To this end,

it is worth to investigate the feasibility of enriching existing lexical resources like

WordNet as well as the possibility of an alternative classification of word senses

based on the intrinsic nature of words, in addition to conventional conceptual

classifications in existing lexical databases.

5 Conclusion

In this study we have analysed word association responses with respect to the lexical

associations obtained from a widely used computational lexicon, namely WordNet. In

particular, we explored the effect of sense abstractness on semantic activation, and

thus the implications on the lexical sensitivity of automatic WSD. Preliminary results

suggest that tangible senses are more readily accessed and syntagmatically related

senses are apparently more strongly associated. The results do not only reinforce the

significance of the intrinsic nature of individual target words in WSD, but also inform

the computational modelling of the internal lexicon and semantic knowledge for the

task.

Acknowledgements

The work described in this paper was fully supported by a grant from the Research

Grants Council of the Hong Kong Special Administrative Region, China (Project No.

CityU 1508/06H).

177

References

1. Collins, A.M. and Loftus, E.F. (1975) A spreading-activation theory of semantic

processing. Psychological Review, 82(6):407-428.

2. Fellbaum, C. (1998) WordNet: An Electronic Lexical Database. MIT Press.

3. Ferret, O. and Zock, M. (2006) Enhancing electronic dictionaries with an index based on

associations. In Proceedings of COLING-ACL 2006, Sydney, Australia, pp.281-288.

4. Hirsh, K.W. and Tree, J.J. (2001) Word association norms for two cohorts of British adults.

Journal of Neurolinguistics, 14:1-44.

5. Hirst, G. (1987) Semantic interpretation and the resolution of ambiguity. Cambridge, UK:

Cambridge University Press.

6. Ide, N. and Veronis, J. (1998) Introduction to the special issue on word sense

disambiguation: The state of the art. Computational Linguistics, 24(1):1-40.

7. Kilgarriff, A. (1992) Polysemy. Ph.D. thesis, University of Sussex.

8. Kwong, O.Y. (2005) Word Sense Classification Based on Information Susceptibility. In A.

Lenci, S. Montemagni and V. Pirrelli (Eds.), Acquisition and Representation of Word

Meaning. Linguistica Computazionale, pp.89-115.

9. Leacock, C., Miller, G.A. and Chodorow, M. (1998) Using Corpus Statistics and WordNet

Relations for Sense Identification. Computational Linguistics, 24(1):147-166.

10. McRoy, S.W. (1992) Using multiple knowledge sources for word sense disambiguation.

Computational Linguistics, 18(1):1-30.

11. Mihalcea, R.F. (2002) Word sense disambiguation with pattern learning and automatic

feature selection. Natural Language Engineering, 8(4):343-358.

12. Mihalcea, R., Chklovski, T. and Kilgarriff, A. (2004) The SENSEVAL-3 English Lexical

Sample Task. In Proceedings of SENSEVAL-3, Barcelona, Spain.

13. Moss, H. and Older, L. (1996) Birkbeck Word Association Norms. Hove, U.K.: Psychology

Press.

14. Quillian, M.R. (1968) Semantic memory. In M. Minsky (Ed.), Semantic Information

Processing. Cambridge, MA: MIT Press.

15. Resnik, P. and Yarowsky, D. (1997) A perspective on word sense disambiguation methods

and their evaluation. In Proceedings of SIGLEX’97 Workshop: Tagging Text with Lexical

Semantics: Why, What, and How?, Washington, D.C., pp.79-86.

16.

Rodd, J., Gaskell, G. and Marslen-Wilson, W. (2002) Making Sense of Semantic

Ambiguity: Semantic Competition in Lexical Access. Journal of Memory and Language,

46:245-266.

17. Swinney, D.A. (1979) Lexical access during sentence comprehension: (Re)consideration of

context effects. Journal of Verbal Learning and Verbal Behavior, 18:645-659.

178