AN E-LIBRARIAN SERVICE THAT YIELDS PERTINENT

RESOURCES FROM A MULTIMEDIA KNOWLEDGE BASE

Serge Linckels

Hasso-Plattner-Institut (HPI), University of Potsdam

Postfach 900460, D-14440 Potsdam

Christoph Meinel

Hasso-Plattner-Institut (HPI), University of Potsdam

Postfach 900460, D-14440 Potsdam

Keywords:

Multimedia, semantic search engine, natural language, information retrieval, performance, e-Learning.

Abstract:

In this paper we present an e-librarian service which is able to retrieve multimedia resources from a knowledge

base in a more efﬁcient way than by browsing through an index or by using a simple keyword search. We

explored the approach to allow the user to formulate a complete question in natural language.

Our background theory is composed of three steps. Firstly, there is the linguistic pre-processing of the user

question. Secondly, there is the semantic interpretation of the user question into a logical and unambiguous

form, i.e. ALC terminology. The focus function resolves ambiguities in the question; it returns the best

interpretation for a given word in the context of the complete user question. Thirdly, there is the generation of

a semantic query, and the retrieval of pertinent documents.

We developed two prototypes: one about computer history (CHESt), and one about fractions in mathematics

(MatES). We report on experiments with these prototypes that conﬁrm the feasibility, the quality and the

beneﬁts of such an e-librarian service. From 229 different user questions, the system returned for 97% of the

questions the right answer, and for nearly half of the questions only one answer, the best one.

1 INTRODUCTION

Our vision is to create an e-librarian service which is

able to retrieve multimedia resources from a knowl-

edge base in a more efﬁcient way than by brows-

ing through an index or by using a simple keyword

search. Our premise is that more pertinent results

would be retrieved if the e-librarian service had a se-

mantic search engine which understood the sense of

the user’s query. This requires that the user must

be given the means to enter semantics. We explored

the approach to allow the user to formulate a com-

plete question in natural langauge (NL). Linguistic

relations within the user’s NL question and a given

context, i.e. an ontology, are used to extract precise

semantics and to generate a semantic query. The e-

librarian service does not return the answer to the

user’s question, but it retrieves the most pertinent doc-

ument(s) in which the user ﬁnds the answer to her/his

question.

The results of our research work are, ﬁrstly, a

founded background theory that improves domain

search engines so that they retrieve fewer but more

pertinent documents. It is based on the semantic in-

terpretation of a complete question that is expressed

in NL, which is to be translated into an unambigu-

ous logical form, i.e. an ALC terminology. Then, a

semantic query is generated and executed. Secondly,

we provide empirical data that prove the feasibility,

and the effectiveness of our underlying background

theory. We developed two prototypes: CHESt (Com-

puter History Expert System) with a knowledge base

about computer history, and MatES (Mathematics Ex-

pert System) with a knowledge base about fractions in

mathematics. We report on experiments with these

prototypes that conﬁrm the feasibility, the quality and

the beneﬁts of such an e-librarian service. From 229

different user questions, the system returned for 97%

of the questions the right answer, and for nearly half

of the questions only one answer, the best one.

In this paper we focus on the translation of a com-

plete NL question into a semantic query. This process

is done in three steps: the linguistic pre-processing

(section 2), the mapping of the question to an on-

tology (section 3), and the generation of a semantic

query (section 4). We present an algorithm (the focus

function) that resolves ambiguities in the user ques-

tion. The outcomes of the experiments are described

in section 5. We present related projects in section 6,

and conclude with some (dis)advantages in section 7.

208

Linckels S. and Meinel C. (2006).

AN E-LIBRARIAN SERVICE THAT YIELDS PERTINENT RESOURCES FROM A MULTIMEDIA KNOWLEDGE BASE.

In Proceedings of the International Conference on Signal Processing and Multimedia Applications, pages 208-215

DOI: 10.5220/0001571302080215

 SciTePress

2 LINGUISTIC

PRE-PROCESSING

The objective of the linguistic pre-processing step

is to convert a stream of symbols into a structured

stream of words, and to retrieve linguistic informa-

tion about these words and the complete sentence.

A search mechanism returns better results if the in-

ference is done over a complete sentence by consid-

ering the relations between words — the syntax —

than by only considering the isolated words. In fact,

the syntactic structure of a sentence indicates the way

words are related to each other, e.g. how the words

are grouped together into phrases, which words mod-

ify which other words, and which words are of central

importance in the sentence.

In our prototypes, the linguistic pre-processing is

performed with a part-of-speech (POS) tagger; we

use TreeTagger (IMS Stuttgart). The linguistic pre-

processing step contributes in three points. Firstly,

the word category of each word is made explicit,

e.g. article, verb. Secondly, the tagger returns the

canonical form (lemma) for each word (token). This

considerably reduces the size of the ontology dic-

tionary. Thirdly, the sentence is split into linguistic

clauses. A linguistic clause is a triple of the form

<subject;verb;object>. Each triple is then processed

individually, e.g. the question q = ”Who invented

the transistor and who founded IBM?” is split into the

two clauses:

′

= [Who invented the transistor?]

conj = [and]

′

= [Who founded IBM?]

3 ONTOLOGY MAPPING

In this section, we present the elaborated background

theory for translating a linguistic pre-processed user

question into a computer readable and unambiguous

form w.r.t. a given ontology.

3.1 Ontology Preliminaries

The e-librarian service masters a domain language

over an alphabet Σ

∗

, which may or may not con-

tain all the possible words L used by the user to for-

mulate his question, so that L

⊆ L ⊆ Σ

∗

. The

semantics are attached to each word by classiﬁcation

in the knowledge source, e.g. a dictionary, which

is structured in a hierarchical way like hyperonym,

hyponym, synonym, and homonyms. In most of the

related projects (section 6), an existing knowledge

Clip

Creator

Person Firm

Software

Language OS

EComponent

Hardware

Computer

Thing

Figure 1: Example of a concept taxonomy about computer

history.

source is used, normally WordNet. The major prob-

lem of such a knowledge source is that it is not ded-

icated to a domain. Like other large scale dictionar-

ies, WordNet on the one hand lacks of speciﬁc do-

main expressions, but on the other hand contains too

much knowledge about other domains. This increases

the problem of ambiguous interpretations for a given

word. We created our own dictionary, which is orga-

nized in a hierarchical way, similar to WordNet, and

w.r.t. our ontology. Furthermore, the size of the dic-

tionary is considerably reduced by the fact that it con-

tains all words from the domain language L

only in

their canonical form. This reduces also the possibility

of ambiguous interpretations.

Deﬁnition 1 (Concept taxonomy) A concept taxon-

omy H = (V, E, v

) is a directed acyclic graph where

each node, except the root-node (v

), has one or more

parents. E is the set of all edges and V is the set of all

nodes (vertices) with V = {(s, T ) | s ∈ S} where s is

a unique label, S the set of all labels in the ontology,

and T is a set of words from L

that are associated

to a node so that T ⊆ L

An example of a concept taxonomy about computer

history is given in ﬁgure 1. Here, a document de-

scribing the transistor would be placed in the concept

”EComponent” (electronic components), which is a

hyponym of ”Hardware”.

A node v

represents a concept. The words that re-

fer to this concept are regrouped in T

. We assume

that each set of words T

is semantically related to

the concept that the node v

represents. The exam-

ple in ﬁgure 2 shows that words like ”Transistor”,

”Diode” or ”LED” semantically refer to the same con-

cept, namely electronic components. Therefore, these

three words are synonyms in the given ontology. Of

course, a certain word can refer to different concepts,

e.g. ”Ada” is the name of a programming language

but also the name of a person. Not all words in L

must be associated with a concept. Only words that

are semantically relevant are classiﬁed. In general,

nouns and verbs are best indicators of the sense of a

question. The difference between words that are se-

mantically irrelevant and words that are not contained

AN E-LIBRARIAN SERVICE THAT YIELDS PERTINENT RESOURCES FROM A MULTIMEDIA KNOWLEDGE

BASE

209

Electronic components

s = EComponent

T = {Transistor, Diode, LED}

Figure 2: Example of a node in the taxonomy about the

concept EComponent (electronic components).

Clip

= ∃hasName.String ⊓ Creator ⊔ T hing

Creator

= P erson ⊔ F irm

P erson

= ∃wasBorn.Date ⊓

∃isDeceased.Date

T hing

= F irm ⊔ Software ⊔ Hardware ⊔ N et

⊓ ∃wasInventedBy.Creator

Sof tware

= Language ⊔ OS

Hardware

= EComponent ⊔ Computer

Figure 3: Example of a concept taxonomy (TBox) about

computer history as ALC terminology.

in L

is that for the second ones, the system has ab-

solutely no idea if they are relevant or not.

3.2 Semantic Interpretation

The representation of context-independent meaning

is called the logical form, and the process of map-

ping a sentence to its logical form is called seman-

tic interpretation (Allen, 1994). The logical form is

expressed in a certain knowledge representation lan-

guage; we use Description Logics (DL). Firstly, DL

have the advantage that they come with well deﬁned

semantics and correct algorithms. Furthermore, the

link between DL and NL has already been established

(Schmidt, 1993). Finally, translating the user question

into DL allows direct reasoning over the OWL-DL en-

coded knowledge base (section 4).

A DL terminology is composed, ﬁrstly, of con-

cepts (unary predicates), which are generally nouns,

question words (w-words) and proper names, and sec-

ondly, of roles (binary predicates), which are gen-

erally verbs, adjectives and adverbs. We use the

language ALC (Schmidt-Schauß and Smolka, 1991),

which is sufﬁciently expressive for our purposes.

ALC concepts are built using a set of concept names

(NC) and role names (NR). Valid concepts (C) are

deﬁned by the following syntax,

C ::= A | ⊤ | ⊥ |

A | C

⊓C

| C

⊔C

| ∀R.C | ∃R.C

with A ∈ NC is a concept name and R ∈ NR is a

role name (ﬁgure 3).

A core part of the semantic interpretation is a map-

ping algorithm. This step — commonly called non-

standard inference (K

usters, 2001) — maps each

word from the user question to one or more ontology

concepts, and resolves the arguments of each role by

analyzing the syntactic structure of the sentence.

Deﬁnition 2 (Word equivalence) The function π :

L, L → R quantiﬁes the similarity of two given words

π(a, b) so that a and b are said to be equivalent w.r.t.

a given tolerance ε, written a ≡ b, iff π(a, b) ≤ ε.

Technically, for a given lemma from the user ques-

tion, the equivalence function π uses the Levenshtein

function to check if this word is contained in the ontol-

ogy dictionary L

given a certain allowed tolerance

ε. That tolerance is calculated relative to the length of

the lemma.

Deﬁnition 3 (Mapping) The meaning of each word

∈ L is made explicit with the mapping function

ϕ : L → V over an ontology dictionary L

⊆ L ⊆

∗

and an ALC concept taxonomy H = (V, E, v

) so

that ϕ(w

) returns a set of interpretations Φ deﬁned

as follows,

Φ = ϕ(w

) = {v

| ∃x ∈ f t(v

) : w

≡ x}.

The function ft(v

) returns the set of words T

as-

sociated to the node v

(deﬁnition 1), and w

≡ x

are two equivalent words. This solution gives good

results even if the user makes spelling errors. Fur-

thermore, only the best matching is considered for the

mapping, e.g. the word ”comXmon” will be consid-

ered as ”common”, and not as ”uncommon”. Both

words, ”common” and ”uncommon”, will be consid-

ered for the mapping of ”comXXmon”. The ambigu-

ity will be resolved in a further step (focus function).

Deﬁnition 4 (Semantic relevance) A word w

is se-

mantically relevant if there is at least one concept in

the ontology H to which w

can be mapped so that

ϕ(w

) 6= ∅.

It is possible that a word can be mapped to different

concepts at once, so that |Φ| > 1. We introduce the

notion of focus to resolve this ambiguity. The focus

is a function (f), which returns the best interpretation

for a given word in the context of the complete user

question.

Deﬁnition 5 (Focus) The focus of a set of interpreta-

tions Φ is made explicit by the function f which re-

turns the best interpretation for a given word in the

context of the complete question q. The focus, written

(ϕ(w

∈ q)) = v

′

, guarantees the following,

1. v

′

∈ ϕ(w

); The focused word is a valid interpre-

tation.

2. |f

(ϕ(w

))| = [0, 1]; The focus function returns 0

or 1 result.

3. ⊤ ≤ v

′

≤ ⊥, if f

(ϕ(w

)) 6= ∅; If the focusing

is successful, then the word is inside the context of

the domain ontology.

SIGMAP 2006 - INTERNATIONAL CONFERENCE ON SIGNAL PROCESSING AND MULTIMEDIA

APPLICATIONS

210

4. π(w

, x ∈ ft(v

′

)) ≤ π(w

, y ∈ ft(v

∈

ϕ(w

))); The returned interpretation contains the

best matching word of all possible interpretations.

Let us consider as illustration the word ”Ada”,

which is called a multiple-sense word. In fact, in the

context of computer history, ”Ada” can refer to the

programming language named ”Ada”, but it can also

be the name of the person ”Augusta Ada Lovelace”.

The correct interpretation can only be retrieved ac-

curately by putting the ambiguous word in the con-

text of a complete question. For example, the context

of the sentences ”Who invented Ada?” and ”Did the

ﬁrms Bull and Honeywell create Ada?” reveals that

here Ada is the programming language, and not the

person Ada.

Technically, the focus function uses the role’s sig-

nature. A role r ∈ NR has the signature r(s

, s

where s

and s

are labels. The signature of each

role deﬁnes the kind of arguments that are possible.

For example wasInventedBy(T hing, Creator) is

the role r = wasInventedBy that has the arguments

= T hing and s

= Creator.

In the question q = ”Who invented Ada?”, ”in-

vented” is mapped to the role w asInventedBy, and

”Who” is mapped to the concept Creator. The sys-

tem detects an ambiguity for the word ”Ada”, which

is mapped to an instance of the concept P erson, but

also to an instance of the concept Language, so that

ϕ(”Ada”) = {P ers on, Language}.

The focus function computes the following combina-

tions to resolve the ambiguity:

1. Was Ada invented by who?*

2. Was Ada invented by Ada?

3. Was who invented by Ada?*

4. Was who invented by who?*

Cyclic combinations like (2) and (4) are not al-

lowed. As for (3), it does not match the role’s sig-

nature because s

= Creator (”Who”), but T hing

is required. As for (1), s

can be P erson or

Language. The role’s signature requires Creator,

therefore Language is excluded as valid interpreta-

tion because Language 6⊑ Creator. As P erson ⊑

Creator, a valid interpretation is found, and in the

context of this question the word ”Ada” refers to the

person Ada. Finally, the result of the focus function

is:

(ϕ(”Ada”)) = P erson.

Indeed, (1) represents the question ”Who invented

Ada?”.

It is still possible that the focus function cannot re-

solve an ambiguity, e.g. a given word has more inter-

pretations but the focus function returns no result:

|Φ| > 1 and f(ϕ(w)) = ∅.

In a such case, the system will generate a semantic

query for each possible interpretation. Based on our

practical experience we know that users generally en-

ter simple questions where the disambiguation is nor-

mally successful.

Deﬁnition 6 (Semantic interpretation) Let q be the

user question, which is composed of linguistic

clauses, written q = {q

′

, ..., q

′

}, with m ≥ 1.

The sematic interpretation of a user question q is the

translation of each linguistic clause into an ALC ter-

minology w.r.t. a given ontology H written,

k=1

′

(ϕ(w

∈ q

′

))

with q

′

a linguistic clause q

′

∈ q, and n the number

of words in the linguistic clause q

′

If a user question is composed of several linguis-

tic clauses, then each one is translated separately.

The logical concatenation of the different interpreted

clauses q

depends on the conjunction word(s) used

in the user question, e.g. ”Who invented the transis-

tor and who founded IBM?”. If no such conjunction

word is found, then the ”or” operator is preferred over

the ”and” operator.

4 QUERY GENERATION

We will start with the assumptions that ﬁrstly, all doc-

uments in the knowledge base K are semantically

described with OWL-DL metadata, w.r.t. an ontol-

ogy H, and that secondly the user question q was

translated into a DL terminology w.r.t. the same on-

tology H (section 3). Even if we currently do not

proﬁt from the full expressivity of OWL-DL, which

is SHOIN (D+), it allows to have compatible se-

mantics between the OWL-DL knowledge base, and

the less expressive ALC user question. Logical in-

ference over the non-empty ABox from K is possible

by using a classical DL reasoner; we use Pellet (Sirin

and Parsia, 2004). The returned results are logical

consequences of the inference rather than of keyword

matchings.

An interpretation I = (∆

, ·

) consists of a non-

empty set ∆

, the domain of the interpretation, and

an interpretation function ·

that maps each concept

name to a subset of ∆

and each role name to a binary

relation r

, subset of ∆

× ∆

Deﬁnition 7 (Semantic query) A semantic query

over a knowledge base K w.r.t. an ontology H, and

an user question q is an ABox query, which means to

search for models I of K, written K |= q

In other words, all documents from the knowledge

base that satisfy the expression q

are potential re-

AN E-LIBRARIAN SERVICE THAT YIELDS PERTINENT RESOURCES FROM A MULTIMEDIA KNOWLEDGE

BASE

211

q = ”Wer hat den Transistor erfunden?”

Linguistic pre-processing

′



Wer

[wer]

WPRO

hat

[haben]

VHFIN

den

[der]

ART

Transistor

[T ransistor]

erfunden

[erf inden]

VVpast



Semantic interpretation

= Creator(x1)

∧

wasInventedBy(x2, x1)

∧

EComponent(x2) ∧ hasName(x2, ”T ransistor”)

Semantic query generation

SELECT ?x1 WHERE (?x2 rdf:type chest:EComponent) (?x2 chest:hasName ?x2hasName)

(?x2 chest:wasInventedBy ?x1) AND (?x2hasName =˜/Transistor/i) USING chest for

<...> rdf for <...>

Figure 4: Complete example for the generation of a semantic query from the user question ”Who invented the transistor?”.

sults. An individual α in I that is an element of (q

)

is a pertinent resource according to the user question.

Technically, an ABox query (in Pellet) is expressed

in a query language; we use RDQL (Miller et al.,

2002) via the Jena framework (Carroll et al., 2004).

Firstly, for a complete question, each semantic inter-

pretation, that is each translated linguistic clause, is

transformed into a semantic query. Secondly, the na-

ture of the question (open or close) reveals the miss-

ing part. An open question contains a question word,

e.g. ”Who invented the transistor?”, whereas a close

question (logical- or yes/no question) does not have a

question word, e.g. ”Did Shockley contribute to the

invention of the transistor?”. As for the ﬁrst kind of

questions, the missing part — normally not an indi-

vidual but a concept — is the subject of the question

and therefore the requested result. The result of the

query is the set of all models I in the knowledge base

K. As for the second kind of questions, there is no

missing part. Therefore, the answer will be ”yes” if

K |= q

, otherwise it is ”no”. A complete example is

shown in ﬁgure 4.

5 IMPLEMENTATION AND

EXPERIMENTS

Our background theory was implemented prototypi-

cally in two educational tools; one about computer

history (CHESt), and one about fractions in mathe-

matics (MatES). Both prototypes can be used at home

or in a classroom either as Web application, or as

stand-alone application (e.g. from a DVD/CD-ROM).

The user can freely formulate a question in NL, and

submit it to the e-librarian service. Then, the e-

librarian service returns one (or more) document(s)

which explain(s) the answer to the user’s question

(ﬁgure 5). The knowledge base is composed of short

multimedia documents (clips), which were recorded

with tele-TASK (http://www.tele-task.de) (Schillings

and Meinel, 2002). Each clip documents one sub-

ject or a part of a subject. The duration of each clip

varies from several seconds to three or four minutes.

This has two reasons, ﬁrstly, the younger the user, the

shorter the time during which (s)he will concentrate

on the information displayed on the screen (Williams

et al., 2001). Secondly, it is not easy to ﬁnd the ap-

propriate information inside a large piece of data, e.g.

in an online lesson that lasts 90 minutes.

In a ﬁrst experiment made in a secondary school

with CHESt, we aimed to investigate, ﬁrstly, how use-

ful our e-librarian service is as an e-learning tool, and

secondly, in how far students accept to enter complete

questions into a search engine instead of only key-

words. Some 60 students took part in the assessment.

In the ﬁrst place, let us point out that nearly all stu-

dents approved of the appealing multimedia presen-

tations. They agreed that the explanations were suf-

ﬁciently complete to understand the subject. Several

appreciated the short length of the clips; a few stated

that the clips were too long. Some added that they

appreciated the short response time of the system. Fi-

nally, asked if they accepted to enter complete ques-

tions into a search engine, 22% of the students an-

swered that they would accept, 69% accepted to enter

complete questions instead of keywords only if this

yielded better results, and 8% disliked this option.

In a second experiment we used MatES to mea-

sure the performance of our semantic search engine.

A testing set of 229 different questions about this

topic was created by a mathematic teacher, who was

not involved in the development of the prototype. The

teacher also indicated manually the best possible clip,

as well as a list of further clips, that should be yielded

as correct answer. The questions were linguistic cor-

rect, and short sentences like students in a secondary

school would ask, e.g. ”How can I simplify a frac-

tion?”, ”What is the sum of

and

?”, ”What are

SIGMAP 2006 - INTERNATIONAL CONFERENCE ON SIGNAL PROCESSING AND MULTIMEDIA

APPLICATIONS

212

Figure 5: CHESt with the question: ’What has Konrad Zuse invented?’.

fractions good for?”, ”Who invented the fractions?”,

etc. This benchmark test was compared with the

performance of a keyword search engine. The key-

word search was slightly optimized to ﬁlter out stop

words (words with no relevance, e.g. articles) from

the textual content of the knowledge base and from

the questions entered. The semantic search engine

answered 97% of the questions (223 out of 229) cor-

rectly, whereas the keyword search engine yielded

only a correct answer (i.e. a pertinent clip) in 70%

of the questions (161 out of 229).

It is also interesting to notice that for 86 questions,

the semantic search engine yielded just one – the se-

mantically best matching – answer (ﬁgure 6). For

75% of the questions (170 out of 229) the semantic

search engine yielded just a few results (one, two or

three answers), whereas the keyword search yielded

for only 14% of the questions less than 4 answers;

mostly (138 questions out of 229) more than 10 an-

swers. Our e-librarian service returned always at least

one result. This is important because we learned from

former experiments in school that students dislike get-

ting no result at all.

For example, the semantic interpretation of the

question ”What is the sum of

and

?” is the fol-

lowing valid ALC terminology:

F raction(x1) ⊓ ∃hasO peration(x1, x2) ⊓

Operation(x2, sum).

Then the semantic query retrieves one clip, which

explained how to add two fractions. This was the best

clip that could be found in the knowledge base

. This

means also that questions like ”How can I add two

fractions”, ”What is

0.5

plus

, etc. would yield the

same clip. The keyword search engine yields all clips,

in which keywords like ”sum” are found, e.g. a clip

that explains how to represent a complex function in

terms of additions, and a clip that explain how to de-

scribe situations with simple fractions.

The experiments revealed also two major weak-

nesses of our e-librarian service that should be im-

proved in future. Firstly, the system is not able to

make the difference between a question, where there

is no answer in the knowledge base, and a ques-

tion that is out of the topic, e.g. ”Who invented

penicillin?”. Secondly, in its current state, the e-

Remember that the system returns clips that explain the

answer to the student’s question, but they do note give the

precise answer, e.g. it does not compute the sum of the two

fractions.

AN E-LIBRARIAN SERVICE THAT YIELDS PERTINENT RESOURCES FROM A MULTIMEDIA KNOWLEDGE

BASE

213

w ith 0

answ er

w ith 1

answ er

w ith 2

answ ers

w ith 3

answ ers

[4..10]

answ ers

[11..20]

answ ers

[21..30]

answ ers

[31..40]

answ ers

> 40

answ ers

100

with 0

ans wer

with 1

ans wer

with 2

ans wers

with 3

ans wers

with 4

ans wers

with 5

ans wers

with 6

ans wers

with 7

ans wers

> 7

ans wers

Figure 6: Number of results yielded by a (1) keyword and by a (2) semantic search engine with a set of 229 questions.

librarian service does not handle number restrictions,

e.g. ”How many machines did Konrad Zuse invent?”.

The response will be the list of Zuse’s machines, but

not a number. Furthermore, the question ”What is the

designation of the third model of Apple computers?”

will yield a list of all models of Apple computers.

6 RELATED WORK

START (Katz, 1997) is the ﬁrst question-answering

system available on the Web. Several improvements

have been made since it came online in 1993 (Katz

and Lin, 2002; Katz et al., 2002) which make of

START a powerful search engine. However, the NLP

is not always sound, e.g. the question ”What did Jodie

Foster before she became an actress?” returns ”I don’t

know what Jodie fostered before the actress became

an actress”. Also, the question ”Who invented the

transistor?” yields two answers: the inventors of the

transistor, but also a description about the transistor

(the answer to the question: ”What is a transistor”).

AquaLog (Lopez et al., 2005) is a portable

question-answering system which takes queries ex-

pressed in NL and an ontology as input, and returns

answers drawn from one or more knowledge bases.

User questions are expressed as triples: <subject,

predicate, object>. If the several translation mech-

anisms fail, then the user is asked for disambiguation.

The system also uses an interesting learning compo-

nent to adapt to the user’s ”jargon”. AquaLog has

currently a very limited knowledge space. In a bench-

mark test over 76 different questions, 37 (48.68%)

where handled correctly.

The prototype PRECISE (Popescu et al., 2003) uses

ontology technologies to map semantically tractable

NL questions to the corresponding SQL query. It was

tested on several hundred questions drawn from user

studies over three benchmark databases. Over 80%

of the questions are semantically tractable questions,

which PRECISE answered correctly, and recognized

the 20% it could not handle, and requests a para-

phrase. The problem of ﬁnding a mapping from the

tokenization to the database requires that all tokens

must be distinct; questions with unknown words are

not semantically tractable and cannot be handled.

FALCON is an answer engine that handles questions

in NL. When the question concept indicating the an-

swer type is identiﬁed, it is mapped into an answer

taxonomy. The top categories are connected to several

word classes from WordNet. Also, FALCON gives a

cached answer if the similar question has already been

asked before; a similarity measure is calculated to see

if the given question is a reformulation of a previous

one. In TREC-9, FALCON generated a score of 58%

for short answers and 76% for long answers, which

was actually the best score.

LASSO relies on a combination of syntactic and se-

mantic techniques, and lightweight abductive infer-

ence to ﬁnd answers. The search for the answer is

based on a form of indexing called paragraph index-

ing. The advantage of processing paragraphs instead

of full documents determines a faster syntactic pars-

ing. The extraction and evaluation of the answer cor-

rectness is based on empirical abduction. A score of

55.5% for short answers and 64.5% for long answers

was achieved in TREC-8.

Medicine is one of the best examples of applica-

tion domains where ontologies have already been de-

ployed at large scale and demonstrated their utility.

The generation, maintenance and evolution of a Se-

mantic Web-based ontology in the context of an in-

formation system for pathology is described in (Bon-

tas et al., 2004). The system combines Semantic Web

and NLP techniques to support a content-based stor-

age and retrieval of medical reports and digital im-

ages.

The MKBEEM (Corcho et al., 2003) mediation

system allows to ﬁll the gap between customers

queries (possibly expressed in NL) and diverse spe-

ciﬁc providers offers. They provide a consensual

representation of the e-commerce ﬁeld allowing the

SIGMAP 2006 - INTERNATIONAL CONFERENCE ON SIGNAL PROCESSING AND MULTIMEDIA

APPLICATIONS

214

exchanges independently of the language of the end

user, the service, or the content provider. The dy-

namic discovery mechanism converts the user query

into an ontological formula, then into a concept de-

scription using DL. Finally, the relevant e-service is

selected. The MKBEEM prototype has been vali-

dated with the languages Finnish, English, French,

and Spanish, in two ﬁelds: business to consumer on-

lines sales, and Web based travel/tourism services.

7 CONCLUSION

In this paper we presented an e-librarian service that

allows the user to communicate by means of complete

questions in NL, and that retrieves pertinent multi-

media resources from a knowledge base. The back-

ground theory is composed of three steps: the linguis-

tic pre-processing of the user’s NL input, the semantic

interpretation of the NL sentence into a logical form,

and the generation of a semantic query. It uses De-

scription Logics and Semantic Web technologies like

OWL for the semantic interpretation of NL questions.

We also presented an algorithm to resolve ambiguities

in the user question. Experiments with two prototypes

conﬁrmed that this background theory is reliable and

can be implemented, e.g. in an educational tool.

In our further work, we will try to improve the

translation from the NL question into an ALC termi-

nology, e.g. use number restrictions. We also want to

investigate if a more precise grammatical analyze of

the user question can help in the interpretation step,

or if this would reduce the users liking of the inter-

face (because of the smaller tolerance of the system).

Another important topic is the maintenance facilities;

how can unknown words from the user query (i.e.

the user’s ”jargon”) be included in the dictionary, and

how can external ”thrusted” knowledge sources been

accessed by the e-librarian service?

REFERENCES

Allen, J. (1994). Natural Language Understanding. Addi-

son Wesley.

Bontas, E. P., Tietz, S., Tolksdorf, R., and Schrader, T.

(2004). Engineering a semantic web for pathology.

In Web Engineering - 4th International Conference,

ICWE 2004, Munich, Germany, pages 585–586.

Carroll, J. J., Dickinson, I., Dollin, C., Reynolds, D.,

Seaborne, A., and Wilkinson, K. (2004). Jena: imple-

menting the semantic web recommendations. In 13th

international conference on World Wide Web - Alter-

nate Track Papers & Posters, New York, NY, USA, May

17-20, pages 74–83.

Corcho,

O., G

omez-P

erez, A., Leger, A., Rey, C., and

Toumani, F. (2003). An ontology-based mediation

architecture for e-commerce applications. In Intel-

ligent Information Processing and Web Mining (IIS:

IIPWM), Zakopane, Poland, pages 477–486.

Katz, B. (1997). Annotating the world wide web using nat-

ural language. In 5th RIAO conference on computer

assisted information searching on the internet, Mon-

treal, Canada.

Katz, B., Felshin, S., Yuret, D., Ibrahim, A., Lin, J. J.,

Marton, G., McFarland, A. J., and Temelkuran, B.

(2002). Omnibase: Uniform access to heterogeneous

data for question answering. In 6th International Con-

ference on Applications of Natural Language to In-

formation Systems (NLDB), Stockholm, Sweden, June

27-28, 2002, Revised Papers, pages 230–234.

Katz, B. and Lin, J. (2002). Annotating the semantic web

using natural language. In 2nd Workshop on NLP and

XML (NLPXML-2002) at COLING.

usters, R. (2001). Non-Standard Inferences in Description

Logics, volume 2100 of Lecture Notes in Artiﬁcial In-

telligence. Springer-Verlag. Ph.D. thesis.

Lopez, V., Pasin, M., and Motta, E. (2005). Aqualog: An

ontology-portable question answering system for the

semantic web. In The Semantic Web: Research and

Applications, Second European Semantic Web Con-

ference, ESWC 2005, Heraklion, Crete, Greece, pages

546–562.

Miller, L., Seaborne, A., and Reggiori, A. (2002). Three

implementations of squishql, a simple rdf query lan-

guage. In 1st International Semantic Web Conference

(ISWC), Sardinia, Italy, June 9-12, pages 423–435.

Popescu, A.-M., Etzioni, O., and Kautz, H. A. (2003). To-

wards a theory of natural language interfaces to data-

bases. In 8th International Conference on Intelli-

gent User Interfaces, January 12-15, Miami, FL, USA,

pages 149–157.

Schillings, V. and Meinel, C. (2002). tele-task: teleteaching

anywhere solution kit. In 30th annual ACM SIGUCCS

conference on User services, Providence, Rhode Is-

land, USA, November 20-23, pages 130–133.

Schmidt, R. A. (1993). Terminological representation, nat-

ural language & relation algebra. In 16th German AI

Conference (GWAI), volume 671 of Lecture Notes in

Artiﬁcial Intelligence, pages 357–371.

Schmidt-Schauß, M. and Smolka, G. (1991). Attributive

concept descriptions with complements. Artiﬁcial In-

telligence, 48(1):1–26.

Sirin, E. and Parsia, B. (2004). Pellet: An owl dl rea-

soner. In International Workshop on Description Log-

ics (DL2004), Whistler, British Columbia, Canada,

June 6-8, volume 104 of CEUR Workshop Proceed-

ings. CEUR-WS.org.

Williams, W. M., Markle, F., Sternberg, R. J., and

Brigockas, M. (2001). Educational Psychology. Allyn

& Bacon.

AN E-LIBRARIAN SERVICE THAT YIELDS PERTINENT RESOURCES FROM A MULTIMEDIA KNOWLEDGE

BASE

215