See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/225163223

Query Expansion with ConceptNet and WordNet: An Intrinsic Comparison

Conference Paper in Lecture Notes in Computer Science · October 2006

DOI: 10.1007/11880592_1 · Source: DBLP

CITATIONS READS

65 3,795

3 authors, including:

Ming-Hung Hsu Ming-Feng Tsai

Industrial technology Research Instittute National Chengchi University
18 PUBLICATIONS 181 CITATIONS 78 PUBLICATIONS 3,837 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Ming-Feng Tsai on 31 May 2014.

The user has requested enhancement of the downloaded file.

 Query Expansion with ConceptNet and WordNet:
An Intrinsic Comparison

Ming-Hung Hsu, Ming-Feng Tsai, and Hsin-Hsi Chen*

Department of Computer Science and Information Engineering

National Taiwan University
Taipei, Taiwan
{mhhsu, mftsai}@nlg.csie.ntu.edu.tw,
hhchen@csie.ntu.edu.tw

Abstract. This paper compares the utilization of ConceptNet and WordNet in

query expansion. Spreading activation selects candidate terms for query expan-
sion from these two resources. Three measures including discrimination ability,
concept diversity, and retrieval performance are used for comparisons. The top-
ics and document collections in the ad hoc track of TREC-6, TREC-7 and
TREC-8 are adopted in the experiments. The results show that ConceptNet and
WordNet are complementary. Queries expanded with WordNet have higher dis-
crimination ability. In contrast, queries expanded with ConceptNet have higher
concept diversity. The performance of queries expanded by selecting the candi-
date terms from ConceptNet and WordNet outperforms that of queries without
expansion, and queries expanded with a single resource.

1 Introduction
Query expansion has been widely used to deal with paraphrase problem in information
retrieval. The expanded terms may come from feedback documents, target document
collection, or outside knowledge resources [1]. WordNet [2], an electronic lexical da-
tabase, has been employed to many applications [9], where query expansion is an im-
portant one. Voorhees [14] utilized lexical semantic relations in WordNet to expand
queries. Smeaton et al [13] added WordNet synonyms of original query terms with half
of their weights. Liu et al [7] used WordNet to disambiguate word senses of query
terms, and then considered the synonyms, the hyponyms, and the words from defini-
tions for possible additions to a query. Roberto and Paola [10] utilized WordNet to ex-
pand a query and suggested that a good expansion strategy is to add those words that
often co-occur with the words of the query. To deal with short queries of web users,
Moldovan and Mihalcea [8] applied WordNet to improve Internet searches.
In contrast to WordNet, commonsense knowledge was only explored in retrieval in
a few papers. Liu and Lieberman [5] used ConceptNet [4] to expand query with the
related concepts. However, the above work did not make formal evaluation, so that
we were not sure if the effects of introducing common sense are positive or negative.
Hsu and Chen [3] introduced commonsense knowledge into IR by expanding con-
cepts in text descriptions of images with spatially related concepts. Experiments
*
Corresponding author.

H.T. Ng et al. (Eds.): AIRS 2006, LNCS 4182, pp. 1 – 13, 2006.
© Springer-Verlag Berlin Heidelberg 2006
2 M.-H. Hsu, M.-F. Tsai, and H.-H. Chen

showed that their approach was more suitable for precision-oriented tasks and for
“difficult” topics. The expansion of this work was done at document level instead of
query level. Document contributes much larger contextual information than query.
In the past, few papers have touched on the comparison of ConceptNet and Word-
Net in query expansion under the same benchmark. We are interested in what effects
of these two resources have. If we know which resource is more useful in a certain
condition, we are able to improve the retrieval performance further. In this paper, we
design some experiments with evaluation criteria to quantitatively measure WordNet
and ConceptNet in the aspect of query expansion. We employ the same algorithm,
spreading activation [12], to select candidate terms from ConceptNet and WordNet
for the TREC topics 301-450, which were used in TREC-6, TREC-7 and TREC-8. To
compare the intrinsic characteristics of these two resources, we propose three types
of quantitative measurements including discrimination ability, concept diversity, and
retrieval performance.
This paper is organized as follows. In Section 2, we give brief introduction to
WordNet and ConceptNet. The comparison methodology is specified in Section 3.
Section 4 introduces the experiment environment and discusses the experimental re-
sults. Section 5 concludes the remarks.

2 WordNet and ConceptNet

In this section, we give brief introduction to the two resources to be compared.
Frameworks and origins of the two knowledgebase are described. A surface compari-
son of their similarities and differences is also presented.

2.1 WordNet

WordNet appeared in 1993 and has been developed by linguistic experts at Princeton
University’s Cognitive Science Laboratory since 1985. It is a general-purpose know-
ledgebase of words, and it covers most English nouns, adjectives, verbs and adverbs.
WordNet’s structure is a relational semantic network. Each node in the network is a
lexical unit that consists of several synonyms, standing for a specific “sense”. Such
lexical unit is called as ‘synset’ in WordNet terminology. Synsets in WordNet are
linked by a small set of semantic relations such as ‘is-a’ hierarchical relations and
‘part-of’ relations. For its simple structure with words at nodes, WordNet’s success
comes from its ease of use [2][9].

2.2 ConceptNet

ConceptNet is developed by MIT Media Laboratory and is presently the largest com-
monsense knowledgebase [6]. ConceptNet is a relational semantic network that is
automatically generated from about 700,000 English sentences of the Open Mind
Common Sense (OMCS) corpus. Nodes in ConceptNet are compound concepts in the
form of natural language fragments (e.g. ‘food’, ‘buy food’, ‘grocery store’, and ‘at
home’). Because the goal of developing ConceptNet is to cover pieces of common-
sense knowledge to describe the real world, there are 20 kinds of relations categorized
as causal, spatial, functional, etc. ConceptNet has been adopted in many interactive
 Query Expansion with ConceptNet and WordNet: An Intrinsic Comparison 3

applications [4]. Hsu and Chen [3] utilized ConceptNet to expand image annotations
and got improvement for some difficult topics. As commonsense knowledge is deeply
context-sensitive, the suitability of ConceptNet for query expansion is still not clear.

2.3 A Surface Comparison

WordNet and ConceptNet have several similarities: (1) their structures are both rela-
tional semantic networks; (2) both of them are general-purpose (that is, not domain-
specific) knowledgebase; and (3) concepts in the two resources are both in the form of
natural language. On the other hand, WordNet and ConceptNet differ from each other
in some aspects: (1) as their processes of development differ (manually handcrafted
vs. automatically generated), intuitively WordNet has higher quality and robustness;
(2) while WordNet focuses on formal taxonomies of words, ConceptNet focuses on a
richer set of semantic relations between compound concepts [6]; and (3) WordNet dif-
ferentiates ambiguous meanings of a word as synsets, however, ConceptNet bears
ambiguity of commonsense knowledge in its concepts and relations.

3 Comparison Methodologies
To compare the two knowledgebase in the aspect of query expansion intrinsically, we
perform the same algorithm to expand queries. As WordNet and ConceptNet are both
relational semantic networks, i.e., useful concepts for expansion in the network are
usually those related to the concepts of the query, spreading activation [12] is adopted.
Figure 1 shows the overall procedure of the comparison. Given an original query,
we perform spreading activation on WordNet and ConceptNet, respectively. Then, the
two expanded queries are compared with three quantitative measurements. The first
measurement computes the discrimination ability in information retrieval. The second
measurement calculates the concept diversity in relevant documents. The third di-
rectly evaluates the performance of retrieval, including two typical evaluation criteria
for ad hoc retrieval. All of these measurements are described in Section 3.4.
When we perform spreading activation in a semantic network to expand a query, the
node of activation origin represents the concept of the given query. The activation ori-
gin is the first to be activated, with an initial activation score (e.g., 1.0). Next, nodes
one link away from the activation origin are activated, then two links away, and so on.
Equation (1) shown below determines the activation score of node j by three fac-
≦
tors: (i) a constant Cdd 1 (e.g., 0.5), which is called distance discount that causes a
node closer to the activation origin to get a higher activation score; (ii) the activation
score of node i; (iii) W(i,j), the weight of the link from i to j. Different relations in the
semantic network are of different weights. Neighbor(j) represents the nodes connected
to node j.
＝ ．
Activation_score( j ) Cdd
i∈Neighbor( j )
．
∑ Activation_score(i) W(i, j ) (1)

Since most traditional IR systems are of bag-of-words model, we select the top N
words with the higher activation scores as the expanded query. For a word w, its acti-
vation score is the sum of scores of the nodes (i.e., synsets in WordNet) that contain w.
4 M.-H. Hsu, M.-F. Tsai, and H.-H. Chen

Fig. 1. Overall procedure in our approach

3.1 Pre-processing

Each query was lemmatized and POS-tagged by Brill tagger. Stop-words were re-
moved and each of the remaining words with POS-tags was considered as a concept
at the following stages.

3.2 Spreading Activation in ConceptNet

In addition to commonsense concepts and relations, ConceptNet also provides a set of

tools for reasoning over text [6]. One of these tools, get_context(concepts), performs
spreading activation on all kinds of relations in ConceptNet, to find contextual
neighborhood relevant to the concepts as parameters. Different relations are set to dif-
ferent weights in default setting. For example, the weight of ‘IsA’ relation is 0.9 and
the weight of ‘DefinedAs’ is 1.0, etc. We adopt this tool directly for collecting ex-
panded terms. In our experiments, each word in a compound concept has the same ac-
tivation score as that of the compound concept. More details about the reasoning tools
in ConceptNet please refer to [6].

3.3 Spreading Activation in WordNet

Since each node in WordNet is a synset that contains synonyms of certain sense,
spreading activation in WordNet is surely performed on the unit of synset. Because
ConceptNet covers most relations in WordNet, we determine the weights of relations
in WordNet, shown in Table 1, by referring to the settings in ConceptNet. For each
concept (a word with POS-tag) in the query, we choose its most frequent sense (syn-
set) as the activation origin from the corresponding POS. In other words, we do not
disambiguate the sense of query terms in this paper for simplicity.
 Query Expansion with ConceptNet and WordNet: An Intrinsic Comparison 5

Table 1. Relations in WordNet and their weights for spreading activation

Relation Type causes holonyms hypernyms hyponyms meronyms Pertainyms

Weight 0.8 0.8 0.8 0.9 0.9 0.7

3.4 Quantitative Measurements

The following proposes three types of measurements to investigate the intrinsic dif-
ferences between WordNet and Concept. They provide different viewpoints for the
comparison.
(1) Discrimination Ability (DA). Discrimination ability is used to measure how
precisely a query describes the information need. In IR, the inverse document fre-
quency (IDF) of a term denotes if it occurs frequently in individual documents but
rarely in the remainder of the collection. For IR systems as the state of the art, dis-
crimination ability of a term can be estimated by its IDF value. Hence the discrimina-
tion ability of a query is measured and contributed by the IDFs of the terms in the
query. For a query q composed of n query terms (q1, q2, …, qn), we define its dis-
crimination ability (DA) as follows.
n

∑ log( df(q ) )
1 NC
DA(q) = (2)
n i =1 i

where NC is the number of documents in a collection, and df(qi) is the document fre-
quency of query term qi.
(2) Concept Diversity (CD). This measurement helps us observe the concept di-
versity of an expanded query, relative to the relevant documents. That is, we measure
how much an expanded query covers the concepts occurring in the relevant docu-
．
ments. Let tm( ) denote the function that maps the parameter (a document or a
query) to the set of its index terms. Let {dq(1), dq(2), …, dq(m)} denote the set of m
documents which are relevant to the query q in the collection. The concept diversity
(CD) of a query q is defined as follows.
m
| tm(q) ∩ tm(d q (i ) ) |
∑
1
CD( q) = (3)
m i =1
| tm(d q (i ) ) |

(3) Retrieval Performance: This type of measurements includes two typical

evaluation criteria for ad hoc retrieval, i.e., average precision (AP) and precision at
top 20 documents (P@20).

4 Experiments and Discussion

We adopted the topics and the document collections in the ad hoc track of TREC-6,
TREC-7 and TREC-8 as the experimental materials for the comparison. There are
556,077 documents in the collection of TREC-6 and 528,155 documents in TREC-7
and in TREC-8. Only the “title” part was used to simulate short query, since web
6 M.-H. Hsu, M.-F. Tsai, and H.-H. Chen

users often submit short queries to search engines. There are totally 150 topics with
identifiers 301-450. However, 4 of them (i.e., topics 312, 379, 392 and 423) are un-
able to be expanded by spreading activation either in WordNet or in ConceptNet, so
that these 4 topics are neglected in the experiments. For each short query, the top 100
words with the higher activation scores form the expanded query. The IR system
adopted for the measurement of retrieval performance is Okapi’s BM25 [11]. The re-
trieval performance is measured on the top 1000 documents for each topic.
Figures 2, 3, 4, and 5 show the results of the quantitative measurements, where the
x-axis represents topic number, and 301-350, 351-400, and 401-450 are topics of
TREC-6, TREC-7 and TREC-8, respectively. To compare the differences between
WordNet and ConceptNet, the result presented for each topic is the difference be-
tween the two expanded queries, i.e., the measurement of the WordNet-expanded
query subtracts that of the ConceptNet-expanded query.

4.1 Preliminary Analysis

Figure 2 shows the differences of two kinds of expansions in discrimination ability

(DA). The DA averaged over the 146 experimental topics is 5.676 and 4.191 for the
WordNet-expanded and ConceptNet-expanded queries, respectively. From Figure 2, it
is obvious that the terms in the WordNet-expanded queries have higher discrimination
ability. In other words, they are more specific than those terms in ConceptNet-
expanded queries. A specific term highly relevant to a topic can be considered as one
of the kernel words of that topic. Figure 2 shows that the queries expanded by Word-
Net are more probable to contain the kernel words.

Difference of DA (W-C)
6

-1

310 320 330 340 350 360 370 380 390 400 410 420 430 440 450
Topic Number

Fig. 2. Differences between WordNet and ConceptNet in discrimination ability (DA)

 Query Expansion with ConceptNet and WordNet: An Intrinsic Comparison 7

0.05
Difference of CD (W-C)
0.04

0.03

0.02

0.01

-0.01

-0.02

-0.03

-0.04

-0.05
310 320 330 340 350 360 370 380 390 400 410 420 430 440 450
Topic Number

Fig. 3. Differences between WordNet and ConceptNet in concept diversity (CD)

0.2 Difference of AP (W-C)

0.1

-0.1

-0.2

-0.3
310 320 330 340 350 360 370 380 390 400 410 420 430 440 450
Topic Number

Fig. 4. Topic-by-topic differences of retrieval performance in average precision (AP)

The average concept diversity (CD) is 0.037 and 0.048 for the WordNet-expanded
and the ConceptNet-expanded queries, respectively. In contrast to the result of dis-
crimination ability, Figure 3 shows that ConceptNet-expanded queries have higher
8 M.-H. Hsu, M.-F. Tsai, and H.-H. Chen

0.5 Difference of P@20 (W-C)

0.4

0.3

0.2

0.1

-0.1

-0.2

-0.3

-0.4

310 320 330 340 350 360 370 380 390 400 410 420 430 440 450
Topic Number

Fig. 5. Topic-by-topic differences of retrieval performance in P@20

concept diversity than WordNet-expanded ones do. Note that the concept diversity of
an expanded query is computed according to the relevant documents. As the terms in
ConceptNet-expanded queries are usually more general than those in WordNet-
expanded queries, Figure 3 shows that ConceptNet-expanded queries cover more of the
concepts that would usually co-occur with the kernel words in the relevant documents.
We call the concepts that co-occur with the kernel words as cooperative concepts.
Expanding a short query with the kernel words will help IR systems to find more
relevant documents. On the other hand, co-occurrence of the cooperative concepts and
the kernel words will help the IR system to rank truly relevant documents higher than
those containing noise. Here we take topic 335 (“adoptive biological parents”) as an
example to illustrate this idea. The kernel words of this topic may be “pregnancy”,
“surrogate”, etc, and the cooperative concepts may be “child”, “bear”, etc. Of these,
“surrogate” and “child” are suggested by WordNet and ConceptNet, respectively. The
pair (child, surrogate) has stronger collocation in the relevant documents than in the
irrelevant documents. The detail will be discussed in Section 4.2.
The overall retrieval performances of the WordNet-expanded queries in AP and
P@20 are 0.016 and 0.0425, respectively. For the ConceptNet-expanded queries, the
performances are 0.019 and 0.0438, in AP and P@20, respectively. These retrieval
performances are low because the expanded queries are formed by the top 100 words
with the higher activation scores. The simple expansion method introduces too much
noise. Figure 4 and Figure 5 show the differences of AP and of P@20 for each topic.
We observed that WordNet-expanded queries perform better for some topics, but
ConceptNet-expanded queries perform better for some other topics. While WordNet
and ConceptNet are different in discrimination ability (Figure 2) and in concept diver-
sity (Figure 3), the two resources can complement each other in the task of ad hoc
retrieval. Hence, we made further experiments in the following subsection.
 Query Expansion with ConceptNet and WordNet: An Intrinsic Comparison 9

4.2 Further Analysis

In the next experiments, we performed manual query expansion by selecting some of

the top 100 words proposed by spreading activation in WordNet or in ConceptNet, to
expand the original query. Two research assistants, each of whom dealt with half of
the topics of TREC-6, performed the process of manual expansion. They read the
topic description in advance, and he/she had no idea about the real content or vocabu-
lary in the relevant documents. This manual selection process was performed sepa-
rately on the words proposed by WordNet and by ConceptNet. These manually se-
lected words for expansion are called WordNet-consulted (WC) and ConceptNet-
consulted (CC) terms, respectively. In this way, we compared four expansion strate-
gies, i.e., original (no expansion), WordNet-consulted, ConceptNet-consulted, and
combination of WC and CC. We also increased the weights of the original query
terms with different degrees to observe how the performances vary with the degrees.
In the experiments, we only used the topics of TREC-6 for analyses.
Figure 6 shows the performances in mean average precision of the four expansion
strategies. The x-axis represents the degrees (times) by which the weights of the
original query terms are increased. The performance of the original query is 0.221 and
doesn’t vary with the degrees since there is no expansion. CC slightly performs better
than the original when the degree is larger than 3, as well as WC with the degree lar-
ger than 6. The slight improvement of CC only or WC only shows that without infor-
mation about the real content of the relevant documents, effective query expansion for
IR is a challenging task even for humans. The best performance is 0.2297, which is
obtained with combination of WC and CC at degree 6, and 3.94% increase to the

0.24

0.22

0.2
MAP

0.18

0.16

0.14 WC+CC
CC
WC
Original
0.12
1 2 3 4 5 6 7 8
Degree

Fig. 6. Performances of the four expansion strategies vs. the weights of the original query term
10 M.-H. Hsu, M.-F. Tsai, and H.-H. Chen

baseline. This performance improvement on no expansion is examined as significant

by a t-test with a confidence level of 95%. The corresponding p-value is 0.034.
Figure 6 also shows that a careful weighting scheme is needed no matter whether
WordNet only or ConceptNet only is adopted. With an unsuitable weighting scheme,
combination of WC and CC performs even worse than WC only or CC only. In Fig-
ure 6, CC performs stably when the degree increases larger than 3, but WC performs
stably only after the degree is larger than 6. As the degree stands for how much the
weights of original query terms are increased, in the aspect of ranking documents, the
degree also stands for how much the weights of CC or WC terms are lightened. While
the words proposed by WordNet are usually more specific and influence more heavily
on the rankings of retrieved documents, it is shown that an appropriate weighting
scheme is more important for WC than for CC.

0.7 Original
WC+CC
CC
0.6 WC

0.5
Average Precision

0.4

0.3

0.2

0.1

0
302 306 307 308 310 319 335 336
Topic Number

Fig. 7. The performances of 8 topics using different expansion strategies

While Figure 6 shows the result averaged over all topics of TREC-6, Figure 7
shows some more strong evidences supporting the argument that WordNet and Con-
ceptNet can complement each other. Using different expansion strategies with the de-
gree 6, the performances (AP) of eight topics are presented. While CC only or WC
only improve the performance of each of the eight topics, it’s obvious that all the
eight topics benefit from the combination of WC and CC. Therefore, the overall im-
provement (refer to Figure 6) of combination of WC and CC is mostly exhibited in
the eight topics.
We verify the complementary of WordNet and ConceptNet, i.e., frequent co-
occurrence of WC terms and CC terms in relevant documents, by the following way.
For each pair of CC term tc and WC term tw, we calculate LRP(tc, tw) using Equation
 Query Expansion with ConceptNet and WordNet: An Intrinsic Comparison 11

(4). LRP(tc, tw) is a value of logarithm of the ratio of two conditional probabilities: (1)
the co-occurrence probability of tc and tw in the relevant documents; and (2) the co-
occurrence probability of tc and tw in the irrelevant documents.
P(tc, tw | R)
LRP(tc, tw) = log (4)
P(tc, tw | IRR)
where R and IRR represent relevant and irrelevant documents, respectively.
Table 2 shows the title, the CC terms, the WC terms and the term pairs having high
LRP values for each topic in Figure 7. The term pairs with high LRP in the eight top-
ics are the major evidences to support that combination of WC and CC is effective as
shown in Figures 6 and 7. Note that the pairs (child, surrogate), (life, surrogate) and
(human, surrogate) of topic 335 have high LRP values. They also confirm the idea of
kernel words and cooperative concepts mentioned in Section 4.1.

Table 2. Illustration of the complementary of WordNet and ConceptNet in 8 topics

Topic Title of Topic CC (tc) WC (tw) (tc,tw): LRP(tc,tw)

302 poliomyelitis and affection disease (global, paralysis): 5.967
post-polio flame place (global, disease): 5.043
global paralysis (affection, paralysis): 4.983
306 african civilian war megadeath (war, killing): 3.715
deaths kill event (kill, casualty): 3.598
nation casualty (nation, killing): 3.408
killing (nation, casualty): 3.357
307 new hydroelectric state proposition (station, risky): 3.906
projects exploration risky (station, examination): 3.598
station examination (exploration, examination): 3.428
308 implant dentistry medical artificial (medical, prosthesis): 9.400
reproductive prosthesis (function, prosthesis): 9.009
function specialty (medical, artificial): 6.980
310 radio waves and cause corpus (cause, radiation): 5.234
brain cancer carcinogen radiation (produce, radiation): 3.958
state (state, radiation): 3.081
produce
319 new fuel sources find material (energy, material): 3.744
energy head (energy, head): 3.217
natural (natural, material): 3.071
335 adoptive biological child surrogate (child, surrogate): 8.241
parents human married (life, surrogate): 6.840
life kinship (human, surrogate): 5.867
336 black bear attacks animal fight (claw, fight): 6.255
claw strike (animal, fight): 3.916
battle counterattack

5 Conclusions and Future Works

In this paper, we used the technique of spreading activation to investigate the
intrinsic characteristics of WordNet and of ConceptNet. Three types of quantitative
12 M.-H. Hsu, M.-F. Tsai, and H.-H. Chen

measurements, i.e., discrimination ability, concept diversity, and retrieval perform-

ance are used to compare the differences between the two resources. With the pre-
liminary analysis and the verification of manual expansion, we have shown that
WordNet is good at proposing kernel words and ConceptNet is useful to find coopera-
tive concepts. With an appropriate weighting scheme, the two resources can comple-
ment each other to improve IR performance.
In future work, we will investigate an automatic query expansion method to com-
bine the advantages of the two resources. Commonsense knowledge in ConceptNet is
deeply context-sensitive so that it needs enough context information for automatic
query expansion. Using existing methods such as pseudo relevance feedback or re-
sources such as WordNet to increase the context information can be explored. We
also intend to investigate whether complex techniques of word sense disambiguation
(WSD) in WordNet are necessary for IR, under the existence of ConceptNet.

Acknowledgements
Research of this paper was partially supported by National Science Council, Taiwan,
under the contracts NSC94-2752-E-001-001-PAE and NSC95-2752-E-001-001-PAE.

References
1. Baeza-Yates, Ricardo and Ribeiro-Neto, Berthier: Modern Information Retrieval, Addi-
son-Wesley (1999).
2. Fellbaum, C. (ed.). WordNet: An Electronic Lexical Database. MIT Press (1998).
3. Hsu, Ming-Hung and Chen, Hsin-Hsi: Information Retrieval with Commonsense Knowl-
edge. In: Proceedings of 29th ACM SIGIR International Conference on Research and De-
velopment in Information Retrieval (2006).
4. Lieberman, H., Liu, H., Singh, P., and Barry, B. Beating Common Sense into Interactive
Applications. AI Magazine 25(4) (2004) 63-76.
5. Liu, H. and Lieberman, H: Robust Photo Retrieval Using World Semantics. In: Proceed-
ings of LREC2002 Workshop: Using Semantics for IR (2002) 15-20.
6. Liu, H. and Singh, P.: ConceptNet: A Practical Commonsense Reasoning Toolkit. BT
Technology Journal 22(4) (2004) 211-226.
7. Liu, S., Liu, F., Yu, C.T., and Meng, W.: An Effective Approach to Document Retrieval
via Utilizing WordNet and Recognizing Phrases. In: Proceedings of the 27th Annual
International ACM SIGIR Conference on Research and Development in Information Re-
trieval (2004) 266-272.
8. Moldovan, D.I. and Mihalcea, R.: Using WordNet and Lexical Operators to Improve
Internet Searches. IEEE Internet Computing 4(1) (2000) 34-43.
9. Morato, Jorge, Marzal, Miguel Ángel, Lloréns, Juan and Moreiro, José.: WordNet Appli-
cations. In: Proceedings of International Conference of the Global WordNet Association
(2004) 270-278.
10. Navigli, Roberto and Velardi, Paola: An Analysis of Ontology-based Query Expansion
Strategies. In: Proceedings of Workshop on Adaptive Text Extraction and Mining at the
14th European Conference on Machine Learning (2003).
 Query Expansion with ConceptNet and WordNet: An Intrinsic Comparison 13

11. Robertson, S.E., Walker, S. and Beaulieu, M.: Okapi at TREC-7: Automatic Ad Hoc, Fil-
tering, VLC and Interactive. In: Proceedings of the Seventh Text Retrieval Conference
(1998) 253-264.
12. Salton, G. and Buckley, C.: On the Use of Spreading Activation Methods in Automatic In-
formation Retrieval. In: Proceedings of the 11th CM-SIGIR Conference on Research and
Development in Information Retrieval (1988) 147-160.
13. Smeaton, Alan F., Kelledy, Fergus, and O'Donell, Ruari: TREC-4 Experiments at Dublin
City University: Thresholding Posting Lists, Query Expansion with WordNet and POS
Tagging of Spanish. In: Proceedings of TREC-4 (1994) 373-390.
14. Voorhees, Ellen M.: Query Expansion Using Lexical-Semantic Relations. In: Proceedings
of the 17th Annual ACM SIGIR Conference on Research and Development in Information
Retrieval (1994) 61-69.

View publication stats