Automatic Extraction of Part-whole Relations

Debela Tesfaye

and Michael Zock

IT doctoral program, Addis Ababa University, Addis Ababa, Ethiopia

Laboratoire d'Informatique Fondamentale (LIF), CNRS, UMR 6166,

Case 901 - 163 Avenue de Luminy, F-13288, Marseille Cedex 9, France

Abstract. The Web has become a universal repository of knowledge allowing

to share information in a scale never seen before. Yet, with the growing size of

the resource have grown the difficulties to access information. In this paper we

present a system that automatically extracts Part-Whole relations among nouns.

The approach is unsupervised, quasi language independent and does not require

a huge resource like WordNet. The results show that the patterns used to extract

Part-Whole relations can be learned from N-grams.

1 Introduction

The Web has become a universal repository of knowledge allowing to share

information in a scale never seen before. Yet, with the expanding size of the resource

have grown the difficulties to access information. Finding the targeted information is

often a tedious and difficult task. Computers still lack too much of the needed

knowledge to understand well the meaning of messages. Hence, information

extraction remains a real challenge.

To overcome this problem researchers have started to develop methods to mine

information from the web. Text mining is knowledge discovery in a corpus via NLP-

and machine-learning techniques. To achieve this goal, i.e. to access the information

contained “inside” the documents, one has to be able to go beyond the information

given explicitly. Put differently, to find a given document one must understand its

content which implies that one is able to identify and understand the concepts and

relations evoked in the document.

There are many kinds of relations, for example: Cause-Effect, Instrument-Agency,

Product-producer, Origine-entity, Theme-tool, Part-whole, Content-container, etc. We

will focus here only on one of them, Part-Whole relations, and their automatic

extraction from corpora. Relations can be used for various tasks: information

extraction, question answering, automatic construction of ontologies, index building

to enhance navigation in electronic dictionaries [17], etc.

The basic premise underlying this work is the idea that the meaning of a word

depends to a large extent on neighborhood be it direct, or indirect: words occurring in

similar contexts tend to have similar meanings [9]. This idea is by no means new. It is

known as the "distributional hypothesis"

and is generally related to scholars like

http://en.wikipedia.org/wiki/Distributional_hypothesis

Tesfaye D. and Zock M..

Automatic Extraction of Part-whole Relations.

DOI: 10.5220/0004113801300139

In Proceedings of the 9th International Workshop on Natural Language Processing and Cognitive Science (NLPCS-2012), pages 130-139

ISBN: 978-989-8565-16-7

 2012 SCITEPRESS (Science and Technology Publications, Lda.)

Harris [8], Firth [3] or Wittgenstein [16]. To understand the meaning of a word

requires that one takes its use, i.e. its context, into account. This latter can be

expressed in terms of (more or less direct) neighbors, i.e. words co-occuring within a

defined window (phrase, sentence, paragraph). One way of doing this is to create a

vector space composed of the target word and its neighbors [11]. We use this kind of

approach to measure the similarity between a pair of words. The vector space model

(VSM) has been developed by Salton and his colleagues [14] for information

retrieval. The idea is to represent each document in a collection as a point in a space,

i.e. a vector in a vector space. Semantic similarity is relative to the distance of two

points: closely related points signal similarity, while distant points signal remotely

related words. We are concerned here with word similarity rather than document

similarity. The meaning of a word is represented as a vector based on the n-gram

value of all co-occuring words. The use of the VSM to extract Part-Whole relations

has two advantages: it requires little human effort and few corpora, at least far less

than Girju's work [4] which relies on annotated corpora and WordNet.

2 Related Work

Several scholars have proposed a taxonomy of Part-Whole relations [4, 14]. We will

follow Winston's [15] classical proposal:

(1) component – integral object handle – cup

(2) member –collection tree – forest

(3) portion –mass grain – salt

(4) stuff – object steel – bike

(5) feature – activity paying – shopping

(6) place – area oasis – desert

Integral objects have a structure; their components can be torn apart and their

elements have a functional relation with respect to the whole. For example, 'kitchen–

apartment' and 'aria–opera' are typical component–integral relations.

'Chairman-committee', 'soldier-army', 'professor-faculty' or 'tree-forest' are typical

representatives of Member-Collection relations.

Portion-Mass captures the relations between portions, masses, objects and

physical dimensions. For example: 'meter-kilometer'.

The Stuff-Object category encodes the relations between an object and the stuff of

which it is made of, be it partly or entirely. For example: 'steel-car', or grape-wine'.

Place-Area captures the relation between an area and a sub-area, or a place within

the area. For example, 'Addis Ababa-Ethiopia'.

Relations can also be categorized according to their frequency. Present part-whole

relations are relations that are always true, while generally absent relations are those

that have happened only at some point in time (episodic presence). Such fine-grained

distinctions are necessary for knowledge representation (acquisition of facts) as well

as for other tasks like question answering, ontologies, etc. Part-whole relations play a

key role in many domains. For example, they are a structuring principle in artifact

design (ships, cars), in chemistry (structure of a substance) and in medicine

131

(anatomy).

Most researchers use supervised learning techniques, relying on numerous features

and large datasets. For example Girju et al. [4] formed a large corpus of 27,963

negative and 29,134 positive examples by using WordNet, the LA Times (TREC9)

and the SemCor 1.7 text to develop three clusters of classification patterns. Matthew

and Charniak [13] applied statistical methods to a very large corpus. This kind of

approach requires manual tagging of the semantic relation of the concepts occuring in

the training set that is expensive in terms of time and use of human resources.

Beamer et al. [1] used the ISA relations, i.e. WordNet's noun hierarchy and the

semantic boundaries built from this hierarchy in order to classify new, i.e. not yet

encountered noun-noun pairs. Unfortunately, the reliance on non-incremental lexical

resources (like WordNet) makes this kind of approach unfeasible for real-world

applications. In addition, it cannot be used for languages devoid of such a resource.

Other approaches are domain dependent. For example, Hage et al. [7] focused on

food ingredients, while Grad's [6] work dealt with bioscience. These approaches are

well suited for a specific domain, but they cannot be used as a general tool.

In addition, the above-mentioned approaches rely on grammar rules (phrase

structure, sentence structure rules, part of speech), which again makes them very

domain- or language dependent, but not well suited as a general tool.

3 Methodology

Our approach is unsupervised, little dependent on language specific and domain

knowledge, and it allows the automatic extraction of meronymic relations

(component–integral object; part-whole). All the system needs is a 'Part of Speech

Tagger' (POS) or a 'part-of-speech tagged corpus' (our case). This is quite different

from other peoples' work in that it does not require a resource like WordNet, which

makes the system useful even for under-resourced language, i.e. languages possibly

lacking a resource like WordNet. The method works for English and it can be

generalized. Hence, it can be adapted to other languages than the one for which it was

initially designed. To achieve this we relied on the 'The Corpus of Historical

American English' (COHA) that contains 400 million words of the period of 1810-

2009 [12]. It is an N-gram corpus tagged for parts of speech [12]. For languages

lacking this kind of tagged corpus, plain text can be used, as the system is able to

identify the concepts' N-gram value in the corpus. This feature is very convenient for

under-resourced languages, making their organization (preparation) easier than

annotating corpora manually. The system contains various components (see 3.1-3.6).

3.1 A POS Tagger

This component identifies the part of speech of the sentence elements. Since part-

whole relations connect only nouns, the system requires only a tagger able to identify

nouns. As mentioned already, we used the 'Corpus of Historical American English'

(COHA) [12]. This is an N-gram corpus whose elements are tagged in terms of part of

speech. The development of a POS tagger being beyond the scope of this paper, we

will not address this issue here.

132

Text

Part-

Whole

Relation

cluster N-N

co-occurences with

identical tail noun

identify N-N

co-occurences

clustering of

head nouns

POS-

tagger

identify similarity

head-noun +

tail noun

identify

part-whole

relationship

Fig. 1. System information flow.

3.2 Identifier of Noun-noun Co-occurrences

This component takes the tagged corpus as input to identify then noun-noun co-

occurrences. There are two types of co-occurrences: nouns occurring directly

together, that is, in adjacent position (NN NN) and nouns whose co-occurrence is

mediated via another type of word occurring in between them (possibly a preposition,

adjective, verbs). Both types need to be identified.

For example, being co-occurrences, the following words are extracted from the

corpus: ‘corolla car’, ‘door of car’, ‘car engine’, ‘engine of car’, ‘car design’,

‘network design’,’ airplane engine’, ‘search engine’ etc. As long as the words linking

two nouns are syntactically different from the linked nouns (verbs, adjectives,

adverbs, preposition), nouns can be easily extracted regardless the type and the

number of words in between them (provided these words are not nouns), even if the

nouns co-occur in a relatively distant location.

At this level we call the noun at the left hand side the head and the one at the right

hand side the tail. 'Car' and 'engine' are respectively the head and the tail in the 'car-

engine' co-occurrence, while they are the reverse in the 'engine-car' co-occurrence.

Hence, cases where the part appears both before and after the whole object will be

retrieved. Since the conclusion that an element is the head or the tail may be

linguistically incorrect we have decided to delay the decision until the end. In other

words, this decision will only be made by the last module (3.6).

This is our first attempt to identify the concepts. The extracted co-occurrences in

this module are the potential concepts. The size of the window of words depends on

the distance between the noun occurrences (the number of words other than noun

occurring between the nouns). Starting from the current noun, we expand the window

size till the next noun occurs. Only concepts inside of the window are considered.

Sentences containing nouns expressing a part whole relationship but whose head

and tail are separated by another noun cannot be retrieved by the system. Example,

‘some tables from the beginning of the 19th century have three legs’. ‘Table’ and

‘legs’ express a part-whole relationship, but this relationship cannot be extracted by

this module as other nouns ‘century/beginning’ occur in between. However, this

relationship can be expressed by a different sentence, contained in the corpus: ‘Some

sophisticated tables have three legs.’ In this case the relationship can be extracted, as

there is no noun in between. This example shows why we should have a well-

balanced corpus, i.e. a corpus rich and representative enough to compensate for

certain cases.

Obviously, the type and number of words other than a noun, yet linking the nouns,

differ from language to language, but this does not have any impact on this module.

133

3.3 Clustering of Noun Co-occurrences with Identical Tail Noun

This module clusters noun co-occurrences on the basis of their tail noun. Noun co-

occurrences sharing the tail noun belong to the same cluster. Hence, 'corolla car' and

'door of car' belong to one cluster, both of them having the same tail noun: 'car', while

'car design' and 'network design' belong to another cluster. The same holds true for

'airplane engine,' 'search engine' and 'car engine'.

car [corolla, door],

design [car, network],

engine [airplane, search, car]

3.4 Clustering of the Head Nouns

This module clusters the head nouns of the noun pairs according to the tail noun. This

clustering is based on the vector similarity value of the head nouns belonging to the

same cluster. The classification is made within the clusters created above. It is based

on the head noun and it is a cluster within a cluster created above. The vector

similarity value is derived from the bi-gram value of a word with all other words in

the corpus (language ?). For this work we used the co-occurrence value (bi-gram) of a

word with all the words occurring in the COCA corpus [12].

All words are represented as a vector of their bi-gram value. Hence, each word has

an N-gram value, represented as a vector. In order to calculate the similarity between

the head nouns we used the cosine value of the vectors of the head noun. Head nouns

whose cosine values are above a certain threshold are clustered together.

Identifying the vector similarity value of words for languages devoid of an N-

gram corpus is very expensive in terms of processing (execution time). Thus, the N-

gram portion of the system will be executed only once, and the result will be stored in

a file to be used later on. This allows using the file at any moment without taxing the

system at the wrong moment.

According to the example given here above 'airplane' and 'car' belong to the same

cluster, while 'search' belongs to another cluster in the 'airplane-engine', 'search-

engine' and 'car-engine' cluster. 'Corolla' and 'door' belong to different clusters in the

'corolla-car' and 'door of car' cluster. Likewise, 'car' and 'network' belong to a different

cluster in 'car-design' and 'network-design'. The clusters are shown below.

car { [corolla] [door] }

design { [car] [network] }

engine {[airplane, car] [search]}

3.5 Identification of the Similarity Value between the Head-noun and the

Tail-noun

This module is the most important one. It identifies the similarity value between the

tail noun and the clusters (head noun). The basic idea is that the tail nouns of the

noun pairs presenting the 'Component-Integral object' or a 'Part-Whole relation' have

a strong similarity value in their clusters.

134

In this module two types of similarity values are calculated. We call them S

and

. The two vectors created for S

and S

in this module are different from the one

used in the module above. The vector for S

is built by considering the words co-

occurring with the tail noun only.

If a word co-occurs both with the tail and the head noun, the N-gram value is

recorded in both vectors, otherwise their respective vector values will be 1 (tail noun)

and zero (head noun). Words co-occuring only with the head noun will not be

included in the vector. Hence, the size of the vector is equal to the size of the number

of words co-occurring with the tail noun. However, in order to create a vector for S

we will also consider words co-occurring with the head noun. The similarity value (C,

for cosinus) between the head nouns in the clusters is calculated on the basis of the bi-

gram value of their co-occurences. Hence, words like 'airplane' and 'car' have a strong

similarity value with respect 'engine', while 'search' has only a small value: airplane-

engine, 'search-engine', 'car-engine'.

3.6 Module for Identifying Part-whole Relations

This module identifies whether two nouns are linked via an integral component Part-

Whole relation or not. To this end, the system relies on information provided by the

above-mentioned modules. From the clusters and the similarity values identified

above in the training corpus, the system extracts automatically a production rule (if

<condition> then <action>) identifying whether two words are linked via an integral

component Part-Whole relation. This rule is very simple, as it requires only a few

steps.

The rules: Given a pair of nouns as described in section 3.4

If the similarity value S

> 0.4 && if the similarity value S

> 0.8

If the noun pairs occurred at least once as compound noun

Then the head noun refers to the whole and the tail to the part

Else

If the average similarity value (C) between the noun and the co-occuring nouns in the cluster >

0.4

If one of the nouns in the cluster has S

> 0.4 and S

> 0.8

If the noun pairs occurred at least once as compound noun

Then the head noun refers to the whole and the tail to the part

Else

The relationship between the nouns is other than a whole-part relation

The rule stipulating that ‘noun pairs occurring at least once as compound noun’, does

not imply that the noun referring to the 'part' is always the second noun, and the

'whole' the first. Indeed, the two may be separated by other types of words, for

example, a preposition, in which case the position of the arguments will swap, the

'part' preceding the 'whole'. Both cases will be handled as discussed in section 3.2.

After extracting the nouns for both cases, we can find the pairs as a compound noun at

least once in a well-balanced corpus. For example, ‘engine of car’ can be extracted as

explained in section 3.2 and the system interpretes the pair as 'part-whole' if it exists

as ‘car engine’, which is always the case in a well-balanced English corpus. For other

languages this should change according to the rules of the langue. Concerning the

135

syntax of part-whole relations it should be noted that, while the two arguments (the

'whole' or the 'part') can in principle be expressed in either order, some languages

allowing even for both of them, every culture, i.e. language, tends to have its own

preferences. For example, in Amharic the order is like in English, in Afan Oromo, the

opposite is the case: the part is mentioned before the whole.

Example:

Cluster 1: Vehicles [car-engine, train-engine, airplane-engine]

Comment: There is an integral component Part-Whole relation, 'engine' being part of

the set of holistic entities: vehicles, car, train, and airplane.

Cluster 2: Oil [benzene-engine, gasoline-engine]

Comment: 'Engine' is not part of 'oil' (benzine or gasoline).

The above two clusters are created within a cluster having engine as supporting ('root'

?) noun. The clusters are identified based on the similarity value among the head

nouns. Accordingly 'car', 'train', and 'airplane' have a strong similarity value. 'Benzine'

and 'gasoline' belong to the same cluster as they have a strong similarity value. The

vector similarity of 'engine' and 'oil' cluster is below the threshold value, while the one

of 'engine' and 'vehicle' is above it. If the similarity valueswere below it, we would

conclude that there is no Part-Whole relationship.

Relational knowledge harvesting is generally based on patterns or clustering. Yet,

it can be improved by vector-based methods. Cederberg and Widdowson [2] showed

how Latent Semantic Analysis (LSA)[10] can help to improve hyponymy extraction

from free text. While LSA is certainly a very powerful method for semantic analysis,

we used a different method. Both approaches rely on word distribution for calculating

semantic similarity, but our approach is different in a number of ways:

1. LSA uses a 'term-by-document-space' to create a vector for a word. This requires

thousands of documents, whereas our approach requires only bi-gram information of

a single well-balanced corpus in order to create a vector of words.

2. LSA uses SVD to reduce the size of the vectors. We do not need SVD at all, as the

vectors are built only on the basis of important N-gram values.

3. LSA considers every word co-occurrence in all documents. In our model the

vector space built for S1 considers only the N-gram value of the tail.

4. LSA uses second order co-occurrence information while we rely only on simple

co-occurrences.

In sum, according to our knowledge, there is no other work trying to extract part-

whole relation via a vector-based approach as we do (including LSA).

4 Evaluation

In order to test our system for the extraction of part–whole relations we used the text

collections of SemEval [4

]. The test corpus is POS-tagged and annotated in terms of

WordNet senses. The corpus has positive and negative semantic relations. The part–

whole relations extracted by the system were validated by comparing them with the

valid relations labeled in the test set answer key. The format of the test set is

described in the sample here below:

136

"Some sophisticated <e2>tables</e2> have three <e1>legs</e1>."

WordNet(e1) = "n3", WordNet(e2)="n2"; Part-Whole(e1, e2) = "true"

This format has been defined by Girju et al [5]. Since this does not correspond to a

real text format, we have changed the corpus accordingly, to obtain the following text:

"Some sophisticated tables have three legs". To evaluate the performance of our

system we deﬁned precision, recall, and F-measure performance metrics in the

following way:

Recall

Number of correctly retrieved relations

Number of correct relations

Precision

Number of correctly retrieved relations

Number of relations retrieved

F-measure

precision recall

Our system identified ALL of the present Component-Integral object part-whole

relation pairs in the SemEval test set. Component-Integral object part-whole relations

in the Semeval training and test set are both present and non-present. This being so,

we considered the present relations to evaluate the performance of our approach. The

performance of the system is evaluated towards identifying present Component-

Integral part whole relation in our test set. As explained in section 3.2 there are

sentences with nouns expressing part whole relations that cannot be detected due to

the fact that there is one or more nouns between the head and the tail. We did not

encounter such sentences in the SemEval test set. However, the nouns can be

extracted from other sentences linking the nouns with words belonging to another

syntactic category. We believe that the nouns can be extracted from a well-balanced

corpus containing other sentences expressing the very same relation by using patterns.

It should be noted that sentence-based evaluation (as there are sentences not

supported by this approach) and corpus-based evaluation yield different results. Of

course, the system should cover both types. To allow for this one could pair all nouns

in a given sentence to form a matrix. The remaining modules will check all the pairs

and those satisfying the constraint specified by rule will be selected.

As the number of present Component-Integral object part-whole relation in the

SemEval test set is small, we added the ones from the SemEval training set and some

of our own in order. The resulting number of relation pairs accounts now for 20% of

our test set. 80 % of this set contains negative examples coming either from the

SemEval test set (all of them) or from our own. As result we achieved a precision of

95, 2%; a recall and F-measure of 95%. All the encountered errors are hyponyms

('car' and 'vehicle'), which does not imply by any means that all the hyponyms in the

test are incorrectly retrieved as part-whole relation. Actually, only 12% of the

hyponyms in the test set are incorrectly retrieved as part-whole relation.

It should also be noted that the majority (80%) of our test set relations are not

present Component-Integral object part-whole relations. Therefore, the probability of

randomly selecting present Component-Integral object part-whole relation is 20/80

(0.25), which shows the effectiveness of this approach for discriminating this kind of

relations.

137

4.1 Comparison with Related Works

As mentionned earlier, other researchers have tried to extract part-whole relation, and

their work [4] generally requires a lot of human resources as it relies on the manual

annotation of the training corpus to extract all the types of the part-whole relations

mentioned in section 2. In addition, it requires the use of WordNet features to extract

the pattern. Also, Girju did not distinguish between 'present part-whole relations' and

'absent' relations as defined by us.

Yet, this kind of distinction is very important for knowledge representation

(acquisition of facts) and question answering. Part-whole relations play a key role in

many applications and in many domains. For example, they are the central structuring

principle in artifact design (ships, cars), in chemistry (structure of a substance) and in

medicine (anatomy). Such relations are omni-present in these domains. Therefore, we

need systems being able to distinguish present part-whole relations from non-present

ones.

Also, most other approaches are highly language dependent. The features used to

build the rules are extracted from a corpus pertaining to a specific language. For

example in Girju et al [4] and Matthew and Charniak [13] most of the rules are based

on prepositions like 'of', 'in', the 's' of the genitive (the

NOUN's) and the auxiliary verb

'to have'. This makes them fairly useless for other languages. Nevertheless, one must

admit that they have covered a wide range of part-whole relation types and that they

did obtain encouraging results.

By contrast, our approach depends little on language: the classification features

are based on word distributions. The features are mainly dependent on the N-gram

value of the concepts, the latter depending entirely on the corpus. The N-gram value

measures the co-occurrence value of the concepts with all other content-bearing

words in the training corpus. This value is represented as a vector allowing to measure

the similarity between a given pair of concepts in a given context. The rules are

learned from word distributions in a corpus only, sentence-structure or phrase-

structure is not considered at all. However, our approach needs to be tested for other

languages in order to attest the validity and the scope of our claims.

5 Conclusions and Future Work

We proposed a new approach to perform semantic analysis. It depends on the

distribution of words for calculating their similarity. The results obtained are quite

promising despite the fact that few resources are used compared to other work. . We

believe that this approach can also be used to extract other types of semantic relations.

The results also showed that the patterns used to extract semantic relations could be

learned from N-gram information.

This latter can be extracted in different ways; we have shown two of them (S

and

) for extracting present Component-Integral object Part-Whole relation. We have

also noted that some hyponyms and present Component-Integral part-whole relation

tend to exhibit to a large extent the same kind of N-gram pattern.

While our current goal has been the identification of part-whole relations, we have

addressed so far only a small part of the problem. Our next step will consist in

extracting the remaining part-whole relations (see, section 2). We also need to check

138

the applicability of our approach for other languages. This should allow us to evaluate

the relative efficiency of the system in language dependency. There are cases where

nouns cannot be extracted by this approach. However, we believe that our approach

can be used to extract present component-integral part-whole relation from a well

balanced-corpus to help building resources like WordNet or an ontology in a given

language. For applications requiring sentence-based semantics, the approach needs to

be modified in order to cover sentences containing nouns with a part-whole relation,

but whose head and tail are separated by one or several nouns.

References

1. Beamer, B., Rozovskaya, A. and Girju, A., (2008). Automatic Semantic Relation Extraction

with Multiple Boundary Generation. Association for the Advancement of Artiﬁcial

Intelligence.

2. Cederberg, S. and D. Widdows (2003). Using LSA and Noun Coordination Information to

Improve the Precision and Recall of Automatic Hyponymy Extraction. In Conference on

Natural Language Learning (CoNLL-2003), Edmonton, Canada, pp. 111-118.

3. Firth, J. R., (1957). A synopsis of linguistic theory 1930-1955. In Studies in Linguistic

Analysis, pp. 1-32. Oxford: Philological Society.

4. Girju R., Moldovan D., Tatu, M. and Antohe, D., (2005). Automatic Discovery of Part–

Whole Relations. ACM 32(1)

5. Girju, R., Hearst, M., Nakov, P., Nastase, V., Szpakowicz; S., Turney, P. and Yuret D.,

(2007). Classification of Semantic Relations between Nominals: Dataset for Task 4 in

SemEval, 4th International Workshop on Semantic Evaluations, Prague, Czech Republic.

6. Grad, B., (1995). Extraction of semantic relations from bioscience text. University of

Trieste, Italy.

7. Hage, W., Kolb, H. and Schreiber, G., (2006). A Method for Learning Part-Whole

Relations. TNO Science & Industry Delft, Vrije Universiteit Amsterdam.

8. Harris, Z., (1954). Distributional structure. Word 10 (23), 46–162.

9. Harshman, R., (1970). Foundations of the parafac procedure: Models and conditions for an

“explanatory” multi-modal factor analysis. UCLA Working Papers in Phonetics, 16.

10. Landauer, T., Foltz P., & Laham, D., (1998). Introduction to Latent Semantic Analysis.

Discourse processes, 25, 259-284.

11. Lund, K. and Burgess, C., (1996). Producing high-dimensional semantic spaces from

lexical co-occurrence. Behavior Research Methods, Instruments, and Computers, 28(2),

203–208.

12. Mark, D., (2011). N-grams and word frequency data from the Corpus of Historical

American English (COHA). http://www.ngrams.info.

13. Matthew, B. and Charniak, E., (1999). Finding parts in very large corpora. In Proceedings

of the 37th Annual Meeting of the Association for Computational Linguistics (ACL1999),

pages 57–64, University of Maryland.

14. Salton, G., Wong, A. and Yang, C.-S., (1975). A vector space model for automatic

indexing. Communications of the ACM, 18(11), 613–620.

15. Winston, M., Chafﬁn, R. and Hermann, D., (1987). Taxonomy of part-whole relations.

Cognitive Science, 11(4), 417–444.

16. Wittgenstein, L., (1922). Tractatus Logico-Philosophicus. London: Routledge & Kegan

Paul.

17. Zock, M.: Needles in a haystack and how to find them? The case of lexical access. In E.

Miyares Bermudez and L. Ruiz Miyares (Eds.) "Linguistics in the Twenty First Century".

Cambridge Scholars Press, 2006, United Kingdom, pp. 155-162.

139