HIERARCHICAL TAXONOMY EXTRACTION BY MINING

TOPICAL QUERY SESSIONS

Miguel Fern

andez-Fern

andez

MVConsultor

ıa, Paseo de la Castellana 91, 28046 Madrid, Spain

Daniel Gayo-Avello

University of Oviedo, Despacho 57, Ediﬁcio de Ciencias, C/Calvo Sotelo s/n 33007 Oviedo, Spain

Keywords:

Web search, Query log, Hyponymy relations, Query reformulation, Automatic taxonomy extraction.

Abstract:

Search engine logs store detailed information on Web users interactions. Thus, as more and more people use

search engines on a daily basis, important trails of users common knowledge are being recorded in those ﬁles.

Previous research has shown that it is possible to extract concept taxonomies from full text documents, while

other scholars have proposed methods to obtain similar queries from query logs. We propose a mixture of both

lines of research, that is, mining query logs not to ﬁnd related queries nor query hierarchies but actual term

taxonomies. In this ﬁrst approach we have researched the feasibility of ﬁnding hyponymy relations between

terms or noun-phrases by exploiting specialization search patterns in topical sessions, obtaining encouraging

preliminary results.

1 INTRODUCTION

Almost half (49%) of the Internet users in the United

States use search engines on a typical day (Fallows,

2008) . This reﬂects the fact that Web search is be-

coming a common habit among users. Because of

this, the amount of data in query logs is steadily in-

creasing every day, thus, recording a great deal of the

common knowledge of the users. As (Pas¸ca, 2007c)

pointed out: “If knowledge is generally prominent or

relevant, people will (eventually) ask about it.”

However, searching is not a straightforward process,

instead, the users gradually reﬁne both their queries

and their goals in a process referred by (Spink et al.,

1998) as the successive search phenomenon. Dur-

ing this iterative process the users provide succes-

sive queries revealing different search patterns (He

et al., 2002). The most relevant ones for this proposal

are the so-called Specialization, Generalization, Re-

formulation, and New. The ﬁrst pattern, Specializa-

tion, refers to the fact that the query q

i+1

deals with

the same topic that q

but seeks more specialized in-

formation (e.g. additional terms have been added to

the query). Generalization refers to the opposite, the

query q

i+1

is on the same topic that q

but seeks more

general information (e.g. some terms have been re-

moved from the original query). In the Reformulation

search pattern both queries are about the same topic

but the user has both added some terms and removed

others from the ﬁrst query and both queries still have

some common terms. The last search pattern, New,

implies that the queries have not any common term

which does not necessarily mean that they are dealing

with different topics.

Although such search patterns just rely on lexical in-

formation (i.e. the presence or absence of terms) we

feel that the number of results satisfying each query

can also provide clues about the existence of special-

ization even when the pattern is just Reformulation

(e.g. dog and labrador, or ipod and electronics).

In addition to that, it must be noticed that when con-

sidering groups of queries we are not interested in

all the queries issued by a user during one “sitting”

(i.e. a searching episode) but in much shorter frag-

ments where all the queries are topically related. The

advantages of using such mini-sessions are two-fold:

(1) the data to be considered in order to ﬁnd seman-

tic relations between terms is much more focused,

and (2) such granularity level should dispel most of

the privacy issues even if no de-identiﬁcation was

used (Xiong and Agichtein, 2007). In order to ob-

tain such query log segmentation we have employed

229

Fernández-Fernández M. and Gayo-Avello D. (2009).

HIERARCHICAL TAXONOMY EXTRACTION BY MINING TOPICAL QUERY SESSIONS.

In Proceedings of the International Conference on Knowledge Discovery and Information Retr ieval, pages 229-235

DOI: 10.5220/0002331402290235

 SciTePress

a technique which has proved to attain similar results

to those achieved by a human expert (Gayo-Avello,

2009). Such technique allows us to group topically

related queries even when those queries do not share

any common term. (see table 1).

Table 1: Five successive records from the AOL query log

grouped into a single topical session by our segmentation

technique. Please notice that from these queries relations

between xenical, xanical (typo), alli, allie (typo) and

fat blocker could be obtained.

Session id Query Clicked Url

6287652 xanical http://trustedmeds.com

6287652 allie or xenical http://stuffonmycat.com

6287652 allie or xenical http://bangornews.com

6287652 alli fat blocker http://700club.com

6287652 alli fat blocker http://wild955.com

2 RELATED WORK

The idea depicted in this paper is somewhat related to

previous and on-going works. We will brieﬂy review

those most relevant and, then, we will point the main

differences between such works and our approach.

First, it must be said that the idea of automatically

building term taxonomies is not new and several ap-

proaches were proposed to work on full text doc-

uments. Works such as (Hearst, 1992) , (Berland

and Charniak, 1999), (Caraballo, 1999), (Girju et al.,

2003), (Morin and Jacquemin, 2003), among others,

are extremely relevant but they cannot be straightfor-

wardly applied to query logs, because most of such

techniques require lexico-syntactic patterns and POS

tagging which are hardly useful when applied to Web

search queries.

With regards to those works relying in query logs (or

in folksonomies) there have been two main goals: (1)

organizing the queries/tags in hierarchical arrange-

ments (but not actual taxonomies), and (2) automat-

ically obtaining similar queries/tags.

Thus, (Clough et al., 2005) and (Schmitz, 2006) ap-

plied subsumption to image tags in order to obtain

tag hierarchies. Such hierarchies, however, were not

taxonomies because no hyponymy relations were es-

tablished; instead, the tags were arranged with re-

gards to their speciﬁcity (e.g. church ← tower

← bell tower, sanfrancisco ← goldengate).

(Heymann and Garcia-Molina, 2006), (Mika, 2007),

and (Schwarzkopf et al., 2007) developed rather sim-

ilar works; they also employed tags collections (al-

though not image tags) and described different tech-

niques to obtain concept hierarchies. Again, such hi-

erarchies were not proper taxonomies.

With regards to the ﬁeld of query suggestion there ex-

ist abundant literature; we will just refer to two re-

cent works that could be confused with our proposal.

For instance, (Shen et al., 2008) and (Baeza-Yates and

Tiberi, 2007) describe two methods to generate re-

lated queries for a given one exploiting the data within

the query log; however, neither of such methods pro-

duces a proper taxonomy in the form we suggest.

Approaches by other authors could be wrongly con-

sidered similar to our approach. For instance,

(Chuang and Chien, 2003) describe a method to clas-

sify query terms into a predeﬁned category system;

thus, it is much closer to query topic classiﬁcation

than to taxonomy bootstrapping. Other works by the

same authors such as (Chuang and Chien, 2004) and

(Chuang and Chien, 2005), describe methods to ob-

tain term hierarchies but such hierarchies are, in fact,

clusters and not taxonomies. There also exist interest-

ing works in the ﬁeld of information extraction. For

instance, (Pas¸ca, 2007a) and (Pas¸ca, 2007c) describe

a technique to obtain class attributes from query logs

(e.g. ﬁnding that population, ﬂag or president are at-

tributes for Country). The same author also provides

a method to ﬁnd named-entities (Pas¸ca, 2007b) which

is related to (Sekine and Suzuki, 2007) and (Komachi

and Suzuki, 2008). None of these works, however, are

related to our approach because they do not generate

term taxonomies.

Thus, our proposal, although somewhat related to all

the aforementioned research is different in several as-

pects. Different from classic works –e.g. (Hearst,

1992), (Berland and Charniak, 1999), (Caraballo,

1999), (Girju et al., 2003) and (Morin and Jacquemin,

2003). in that it does not rely in full text docu-

ments but in query logs. It also differs from (Clough

et al., 2005), (Heymann and Garcia-Molina, 2006),

(Schmitz, 2006), (Mika, 2007), (Baeza-Yates and

Tiberi, 2007) and (Schwarzkopf et al., 2007) in the

underlying goal: while those methods obtain tags or

queries hierarchies according to their speciﬁcity, we

are interested in automatically building actual tax-

onomies (i.e. hierarchical arrangements according to

hyponymy relations). We have also exposed that other

works such as (Chuang and Chien, 2003), (Chuang

and Chien, 2004), (Chuang and Chien, 2005), (Pas¸ca,

2007c), (Pas¸ca, 2007a), (Pas¸ca, 2007b), (Sekine and

Suzuki, 2007) and (Komachi and Suzuki, 2008) are in

fact dealing with problems totally unrelated to taxon-

omy construction.

KDIR 2009 - International Conference on Knowledge Discovery and Information Retrieval

230

3 MOTIVATION

By using taxonomies it should be possible to greatly

improve search engine results by means of term dis-

ambiguation, query suggestion and expansion. How-

ever, we feel that current lexical databases (e.g. Word-

Net (Miller, 1990)) have several issues in order to be

really useful for such purposes. First, because Word-

Net is an English language project, parallel projects

for other languages have been developed, such as Eu-

roWordNet (Vossen, 1998), BalkaNet (Greek), He-

brew WordNet, Hindi WordNet and Japanese Word-

Net among others (Vossen and Fellbaum, 2004). Cer-

tainly we could rely on such different wordnets but

the task of identifying the language in which queries

are written is not trivial given the small number of

terms usually employed. Additionally, there exist

a huge gap between the lexicon used by Web users

and the developers of wordnets. For instance, (Man-

dala et al., 1999) and (Gabrilovich and Markovitch,

2007) pointed out that most domain-speciﬁc relation-

ships between words cannot be found in WordNet,

and some kind of words, such as proper names, jar-

gon or slang are just not included. Besides, (Mi-

halcea, 2003) also explained that due to the fact

that professional linguists recognize minimal differ-

ences in word senses, common words such as “make”

have too many different senses to be useful for IR

tasks. Of course, such wordnets could be auto-

matically enriched (Hearst, 1992) but such approach

require a great effort (usually carried out by lin-

guists) and, hence, wordnets remain as quite static

data sources. On the other hand, most of the afore-

mentioned methods to construct term taxonomies –

e.g. (Hearst, 1992), (Berland and Charniak, 1999),

(Caraballo, 1999), (Girju et al., 2003), (Morin and

Jacquemin, 2003). not only need large text corpora

but they are tightly coupled to the grammar rules of

the target language. This would make their appli-

cation to query logs extremely difﬁcult (if not to-

tally unfeasible) given the nature of the queries which

are short and, many times, simply ungrammatical.

Thus, we feel that taxonomies of terms and noun

phrases collecting the common knowledge of search

engine users, including typos, jargon and slang are

a real need in order to improve the performance of

Web search engines. Besides, we think that the only

way to obtain such users’ mental model is by min-

ing the query logs collecting the users queries. As a

consequence, the following research questions arose:

(1) Is it possible to generate term taxonomies rely-

ing only on the queries submitted by search engine

users? And, (2) Can they be automatically mined in a

language-independent way?

Throughout the following sections we describe

our proposal to mine hyponymy relations from query

logs, we provide preliminary results from its applica-

tion to the AOL dataset, in addition to future lines of

research.

4 RESEARCH DESIGN

4.1 The AOL dataset

To answer the research questions stated above we em-

ployed the AOL query log (Pass et al., 2006). This

data set contains more than 30 million records from

about 650,000 users sampled from March to May

2006. Each of those records comprises (1) a user iden-

tiﬁer, (2) the query string submitted by the user, and

(3) the date and time when the query was issued. If the

user clicked any result, then the record also includes

the host name portion of the clicked URL. As it was

previously explained, in addition to that information,

we also need the estimated number of results for each

query to test the existence of subsumption. Although

such information could appear in query logs (e.g. the

MSN query log (Microsoft, 2006) does contain it) it

is not the case with the AOL data and, hence, we had

to enrich the original information by means of the Ya-

hoo! BOSS API

4.2 Data Preparation

We preprocessed the raw log in order to obtain a sub-

set best suited to our purposes. In this phase we ses-

sionized the log and removed those records which

were supposed to tamper with the extraction phase.

Because of its relevance, that phase will be detailed

in a separate section. With regards to the preprocess-

ing phase it comprises the following steps:

1. Removal of Navigational Queries. According

to (Broder, 2002), there are three kind of queries

in relation to their intent: (1)Navigational, when

the immediate intent is to reach a particular site;

(2) Informational, when the intent is to acquire

some information assumed to be present on one

or more web pages; and (3)Transactional, when

the intent is to perform some web-mediated ac-

tivity (e.g. to buy a product, or download a ﬁle).

We think that navigational queries can reduce the

accuracy of the hyponym extraction process and,

thus, such queries should be removed. Because

the intent behind navigational queries is to reach a

particular site, most of them are lexically similar

http://developer.yahoo.com/search/boss/

HIERARCHICAL TAXONOMY EXTRACTION BY MINING TOPICAL QUERY SESSIONS

231

to the URL of the referred site and, thus, a sim-

ple heuristic to detect them (Jansen et al., 2008)

consists of checking for the appearance of well-

known website names (e.g. google, wikipedia,

etc), domain sufﬁxes (e.g. com, net, ... , co.uk,

etc), or strings frequently present in URLs such as

www. or http://.

2. Log Sessionization. Topical query sessions are

sequences of queries issued by a single user and

dealing with a unique topic. There are several

methods well suited for this task and for these

experiments we have applied the technique that,

according to the analysis performed in (Gayo-

Avello, 2009), achieves the best results. Such a

technique relies in both lexical and temporal clues

to ﬁnd a topic shift or a topic continuation be-

tween successive queries from the same user.

3. Removal of Repeated Queries. Many sessions

contain records with repeated queries; there are

several reasons for this: maybe the user clicked

on more than one search result after submitting

the query; s/he could also ask for another page of

results; or even the user actually typed the same

query more than once during the same session.

For our purposes, more than one record with the

same value for the query ﬁeld is simply redundant

and, hence, repeated records were removed.

4. Removal of Low Frequency Queries. Nonsen-

sical queries or containing odd typos are not rare

but they are relatively unfrequent. In order to miti-

gate the impact of such queries, we removed those

records containing queries with an absolute fre-

quency below an arbitrary threshold.

5. Removal of One-query-sessions. Because our

proposal to detect hyponym relations requires a

specialization pattern, sessions containing just

one query are useless and, therefore, removed.

4.3 Taxonomy Extraction

4.3.1 Identiﬁcation of Specializations

Our proposal is based on the premise that hyponymy

relations can be mined from query logs by taking into

account the specialization search pattern in addition

to the number of results available for each query.

Specialization occurs when a query q

looks for in-

formation on the same topic than a previous query

but it is much more focused. To detect such a

pattern researchers have typically relied in lexical

similarity [(Miller, 1990), (He et al., 2002)], that

is, a query is an specialization when it adds new

terms to a previous query. A trivial scenario oc-

curs when the query q

contains as a substring the

previous query q

(e.g. fish food and tropical

fish food); we will refer to this as trivial special-

ization. A not so trivial case occurs when the subse-

quent query not only add new terms but remove others

from the previous query, such as in the pair angelica

panganiban scandals and celebrity scandals.

Arguably, this example is considered an specializa-

tion just because the number of terms is larger but,

in fact, angelica panganiban is a multiword term

and, thus, such a pair could also be considered a case

of Reformulation. In fact, apart from the appearance

of a multiword term, that case would be analogous to

the pair comprising electronic repairs and ipod

repairs that clearly ﬁts the Reformulation pattern.

It is because of such cases that we propose to use

the number of results obtained by each query; that

way, if the number of results is signiﬁcantly differ-

ent we could consider such reformulations as refor-

mulations with specialization, even, if they do not add

new terms. Besides, it must be noticed that there

exist many queries that do not ﬁt within those in-

terpretations and are, however, specializations (e.g.

labrador and dog). Nevertheless, at this early stage

we are only considering trivial specializations and re-

formulations with specialization. Finally, to increase

recall, we have taken every combination of queries

within the same topical session instead of just the

pairs of successive queries. From such pairs, depend-

ing on the type of specialization detected (if any), two

different extraction algorithms are applied.

4.3.2 Extraction of Hyponymy Relations

When a reformulation with specialization occurs, we

check the number of results obtained by each query in

the pair. If such numbers are signiﬁcantly different we

assume that the query with a larger number of results

subsumes the other one. Then, the hyponym and hy-

pernym are obtained by removing the terms appearing

in both queries (see Figure 1). It must be noticed that

this process is not error-free and some noise is, at this

moment of the research, unavoidable; for instance,

inferring from the queries pet hamsters (623K re-

sults) and pet dogs (4M results) that hamsters spe-

cializes dog; such issues must be addressed in future

research.

When dealing with trivial specializations, the pre-

vious method is unfeasible because as one of the

queries is a substring of the other when subtract-

ing the intersection between both queries from the

shorter one it would result in the empty set. We can,

of course, directly take the pair as a hyponym re-

lation (e.g. fish food ← tropical fish food),

KDIR 2009 - International Conference on Knowledge Discovery and Information Retrieval

232

Figure 1: Extraction of a relation from a reformulation with

specialization pair. In this case, we would infer that “trade

edition” is a hyponym of “book” which is indeed correct (a

trade edition is a book intented for general readership)

but other hyponymy relations could also be inferred

from the same pattern, such as fish ← tropical

fish, and food ← tropical fish food. To ex-

tract such relations, we employ a method relying

on the use of term n-grams. Firstly, we produce

for each of the queries every possible n-gram (e.g.

for the query tropical fish food we would ob-

tain tropical, fish, food, tropical fish, fish

food, and tropical fish food). Then, every n-

gram of the specialized query is paired with every n-

gram of the generalized query provided that (1) both

n-grams contain common terms, (2) they are not the

same n-gram, and (3) the n-gram from the special-

ized query is longer than the n-gram from the gener-

alized query. Continuing with the example, the pairs

ﬁsh food and food, or tropical ﬁsh and ﬁsh, would be

obtained, but also ﬁsh food and ﬁsh (see Figure 2)

fish ← tropical fish

fish ← fish food

food ← fish food

fish ← tropical fish food

food ← tropical fish food

fish food ← tropical fish food

Figure 2: List of candidate hyponymy relations obtained

from the trivial specialization pair (tropical ﬁsh food, ﬁsh

food). As it can be seen, the second and fourth candidates

are incorrect.

5 PRELIMINARY RESULTS

We have applied the methods described above to

a sessionized and preprocessed version of the AOL

query log obtaining encouraging results. At this mo-

ment we have not yet developed a way to cuantita-

tively evaluate the accuracy of the results and, hence,

we can only provide a short sample of the relations

obtained that we consider illustrative of the pros and

cons of our technique:

• coin ← penny and military ← navy. These

relations are not only correct, but they also appear

in Wordnet.

• lingerie ← panties. Another relationship

that should be considered correct although it does

not appear in such a straightforward way in Word-

net.

• celtic ← irish. This relation is highly rea-

sonable and reveals many of the problems of hi-

erarchical taxonomies; that is, are we referring to

the Irish language (which indeed belongs to the

Celtic family) or to the Irish people?

• eventing ← jumping. This relation is one the

most frequent, appearing in over 300 different ses-

sions. It illustrates very well the way in which

speciﬁc domains can be exposed because event-

ing is an equestrian competion comprising several

disciplines including jumping; such sense does

not appear in Wordnet.

• motels ← howard johnson express and

wrestling ← wwe. These relations are correct

because Howard Johnson Express and WWE

are brands dealing with motels and professional

wrestling respectively. Because they involve

trademarks, they do not appear in Wordnet.

• underwear ← briefs ← speedo. Here we

have two hyponymy relations: underwear ←

briefs and briefs ← speedo. The ﬁrst re-

lation is correct because briefs are a type of

tight underwear and swimwear. The second re-

lation could also be considered correct because,

although Speedo is a trademark, it is commonly

used as a generic name referring to swimming

briefs. Again, these relations reveal common

knowledge that is not usually present in lexical

databases.

• mountain ← mountian, paper ← papper, or

video ← vidio are examples of rather frequent

relations mined from the query log. As it can be

seen they are not really hyponyms but typos asso-

ciated with the correct spelling.

• hanoverian ← arabian, yellow ← white,

honda ← kawasaki, justice leage ←

flash gordon are examples of some of the

issues we have to face up in future research; in

these cases we have terms that could be consid-

ered co-hyponyms (i.e. terms with a common

hypernym) but one of them is much more popular

than the other, thus, tampering with our heuristic

based on the number of results.

HIERARCHICAL TAXONOMY EXTRACTION BY MINING TOPICAL QUERY SESSIONS

233

6 CONCLUSIONS AND FUTURE

WORK

As it has been exposed above, lexical databases are

costly hand made resources that, however, exhibit a

lack of common dayly knowledge such as jargon,

slang and frequent typos. Nevertheles, such terms, be-

cause of their pervasive presence in user Web search

queries, are extremely important to improve the per-

formance of search engines. This fact drove us to re-

search the feasibility of automatically extracting term

taxonomies from those very same queries. Along this

paper we have described an approach with encourag-

ing preliminary results. In fact, it seems that it is not

only possible to achieve such results by only using

query logs but also that it should be possible to at-

tain that in different languages. Therefore, the stated

research questions seem to have a feasible answer

This research also has limitations that should be ad-

dressed in the near future. First, a much more pre-

cise way to identify specialization patterns is needed.

Second, false positives (i.e. incorrectly ﬂagged hy-

ponymy relations) should be ﬁltered out. And third,

an evaluation framework should be envisioned in or-

der to quantify the performance of the method. With

regards to the ﬁrst issue we have also pointed out that,

at this moment, only lexical clues are employed to de-

tect specialization but we plan to reproduce the work

by (Boldi et al., 2009) where they describe a machine

learning method to detect much subtle specializations

(e.g. labrador and dog). Regarding the second is-

sue, we have explored a na

ıve heuristic based on the

position where modiﬁers occur in relation to the hy-

pernym (i.e. they are pre- or post-modiﬁers). As we

pointed before, in the English language such modi-

ﬁers tend to precede the hypernym (e.g. tropical

fish, blue fish, recently caught fish) and,

hence, it could be rather simple to remove most of

the false positives. This could work in other lan-

guages but, certainly, it would not be language inde-

pendent. However, statistical methods could perhaps

be applied to these trivial specializations to discover

the most common position of modiﬁers in order to

adapt the application of the heuristic. Finally, with re-

gards to the third issue on the necessity of a evaluation

framework, we will probably start relying on Word-

net although we have already pointed out the lack of

specialized knowledge and slang in that database. On

the other hand, we believe that many pairs would be,

in fact, instances and not hyponyms (e.g. angelina

jolie and celebrity) which could be really difﬁ-

cult to evaluate by simply using Wordnet. Hopefully,

in future works we will be able to shed light on such

issues.

REFERENCES

Baeza-Yates, R. and Tiberi, A. (2007). Extracting seman-

tic relations from query logs. In KDD ’07: Proc. of

the 13th ACM SIGKDD international conference on

Knowledge discovery and data mining, pages 76–85,

New York, NY, USA. ACM.

Berland, M. and Charniak, E. (1999). Finding parts in very

large corpora. In Proceedings of the 37th annual meet-

ing of the Association for Computational Linguistics

on Computational Linguistics, pages 57–64.

Boldi, P., Bonchi, F., Castillo, C., Donato, D., and Vi-

gna, S. (2009). Query suggestions using query-ﬂow

graphs. In WSCD ’09: Proc. of the 2009 workshop on

Web Search Click Data, pages 56–63, New York, NY,

USA. ACM.

Broder, A. (2002). A taxonomy of web search. SIGIR Fo-

rum, 36(2):3–10.

Caraballo, S. A. (1999). Automatic construction of a

hypernym-labeled noun hierarchy from text. In Pro-

ceedings of the 37th annual meeting of the Association

for Computational Linguistics on Computational Lin-

guistics, pages 120–126, Morristown, NJ, USA. As-

sociation for Computational Linguistics.

Chuang, S.-L. and Chien, L.-F. (2003). Enriching web

taxonomies through subject categorization of query

terms from search engine logs. Decis. Support Syst.,

35(1):113–127.

Chuang, S.-L. and Chien, L.-F. (2004). A practical web-

based approach to generating topic hierarchy for text

segments. In CIKM ’04: Proc. of the thirteenth ACM

international conference on Information and knowl-

edge management, pages 127–136, New York, NY,

USA. ACM.

Chuang, S.-L. and Chien, L.-F. (2005). Taxonomy gener-

ation for text segments: A practical web-based ap-

proach. ACM Trans. Inf. Syst., 23(4):363–396.

Clough, P., Joho, H., and Sanderson, M. (2005). Automati-

cally organising images using concept hierarchies,. In

Proc. of the SIGIR Workshop on Multimedia Informa-

tion Retrieval.

Fallows, D. (2008). Almost half of all internet users

now use search engines on a typical day. Tech-

nical report, Pew Internet & American Life

Project. Accessed 6 February 2009. Available at:

http://www.pewinternet.org/pdfs//PIP Search Aug08.pdf.

Gabrilovich, E. and Markovitch, S. (2007). Harnessing the

expertise of 70,000 human editors: Knowledge-based

feature generation for text categorization. J. Mach.

Learn. Res., 8:2297–2345.

Gayo-Avello, D. (2009). A survey on session detection

methods in query logs and a proposal for future eval-

uation. Inf. Sci., 179(12):1822–1843.

Girju, R., Badulescu, A., and Moldovan, D. (2003). Learn-

ing semantic constraints for the automatic discovery

of part-whole relations. In NAACL ’03: Proc. of the

2003 Conference of the North American Chapter of

the Association for Computational Linguistics on Hu-

man Language Technology, pages 1–8, Morristown,

NJ, USA. Association for Computational Linguistics.

KDIR 2009 - International Conference on Knowledge Discovery and Information Retrieval

234

He, D., G

oker, A., and Harper, D. J. (2002). Combining

evidence for automatic web session identiﬁcation. Inf.

Process. Manage., 38(5):727–742.

Hearst, M. A. (1992). Automatic acquisition of hyponyms

from large text corpora. In Proceedings of the 14th

conference on Computational linguistics, pages 539–

545, Morristown, NJ, USA. Association for Compu-

tational Linguistics.

Heymann, P. and Garcia-Molina, H. (2006). Collaborative

creation of communal hierarchical taxonomies in so-

cial tagging systems. Technical Report 2006-10, Stan-

ford University.

Jansen, B. J., Booth, D. L., and Spink, A. (2008). Deter-

mining the informational, navigational, and transac-

tional intent of web queries. Inf. Process. Manage.,

44(3):1251–1266.

Komachi, M. and Suzuki, H. (2008). Minimally supervised

learning of semantic knowledge from query logs. In

Proc. of the 3rd International Joint Conference on

Natural Language Processing (IJCNLP 2008), pages

358–365.

Mandala, R., Tokunaga, T., and Tanaka, H. (1999). Comple-

menting wordnet with roget’s and corpus-based the-

sauri for information retrieval. In Proceedings of the

ninth conference on European chapter of the Associ-

ation for Computational Linguistics, pages 94–101,

Morristown, NJ, USA. Association for Computational

Linguistics.

Microsoft (2006). Microsoft Research Microsoft Live

Labs: Accelerating Search in Academic Research

2006. Available at: http://research.microsoft.com/

ur/us/fundingopps/RFPs/Search

2006 RFP.aspx. (Ac-

cessed 24 November 2008).

Mihalcea, R. (2003). Turning wordnet into an informa-

tion retrieval resource: Systematic polysemy and con-

version to hierarchical codes. International Jour-

nal of Pattern Recognition and Articial Intelligence

(IJPRAI), pages 689–704.

Mika, P. (2007). Ontologies are us: A uniﬁed model of

social networks and semantics. Web Semant., 5(1):5–

15.

Miller, G. A. (1990). Wordnet: An on-line lexical database.

International Journal of Lexicography, pages 235–

312.

Morin, E. and Jacquemin, C. (2003). Automatic acquisition

and expansion of hypernym links. Computer and the

humanities, 38:363–396.

Pas¸ca, M. (2007a). Organizing and searching the world

wide web of facts – step two: harnessing the wisdom

of the crowds. In WWW ’07: Proc. of the 16th interna-

tional conference on World Wide Web, pages 101–110,

New York, NY, USA. ACM.

Pas¸ca, M. (2007b). Weakly-supervised discovery of named

entities using web search queries. In CIKM ’07: Proc.

of the sixteenth ACM conference on Conference on

information and knowledge management, pages 683–

690, New York, NY, USA. ACM.

Pas¸ca, V. D. (2007c). What you seek is what you get: Ex-

traction of class attributes from query logs. In Pro-

ceedings of the 20th International Joint Conference on

Artiﬁcial Intelligence (IJCAI-07), pages 2832–2837.

Pass, G., Chowdhury, A., and Torgeson, C. (2006). A pic-

ture of search. In InfoScale ’06: Proc. of the 1st inter-

national conference on Scalable information systems,

page 1, New York, NY, USA. ACM.

Schmitz, P. (2006). Inducing ontology from ﬂickr tags.

In Proc. of the Collaborative Web Tagging Workshop

(WWW ’06).

Schwarzkopf, E., Heckmann, D., Dengler, D., and Kro-

ner, A. (2007). Mining the structure of tag spaces for

user modeling. In Workshop on Data Mining for User

Modeling.

Sekine, S. and Suzuki, H. (2007). Acquiring ontological

knowledge from query logs. In WWW ’07: Proc. of

the 16th international conference on World Wide Web,

pages 1223–1224, New York, NY, USA. ACM.

Shen, D., Qin, M., Chen, W., Yang, Q., and Chen, Z. (2008).

Mining web query hierarchies from clickthrough data.

In AAAI07: Proc. of the Twenty-Second Conference

on Artiﬁcial Intelligence.

Spink, A., Wilson, T., Ellis, D., and Ford, N. (1998). Model-

ing users’ successive searches in digital environments.

D-Lib Magazine. Accesed 6 February 2009. Available

at: http://www.dlib.org/dlib/april98/04spink.html.

Vossen, P., editor (1998). EuroWordNet: a multilingual

database with lexical semantic networks. Kluwer

Academic Publishers, Norwell, MA, USA.

Vossen, P. and Fellbaum, C. (2004). Wordnets in the

world. Technical report, Global WordNet Association

[http://www.globalwordnet.org/]. Accessed 06-02-09.

Xiong, L. and Agichtein, E. (2007). Towards privacy-

preserving query log publishing’. In Query Log Anal-

ysis: Social And Technological Challenges. A work-

shop at the 16th International World Wide Web Con-

ference (WWW2007).

HIERARCHICAL TAXONOMY EXTRACTION BY MINING TOPICAL QUERY SESSIONS

235