GOSSIP GALORE

A Conversational Web Agent for Collecting and Sharing Pop Trivia

Feiyu Xu, Peter Adolphs, Hans Uszkoreit, Xiwen Cheng and Hong Li

DFKI GmbH, Language Technology Lab, Stuhlsatzenhausweg 3, D-66123 Saarbr¨ucken, Germany

Keywords:

Web mining, Relation extraction, Web intelligence, Intelligent user interface, Conversational agent, Question

answering, Dialogue system.

Abstract:

This paper presents a novel approach to a self-learning agent who collects and learns new knowledge from

the web and exchanges her knowledge via dialogues with the users. The application domain is gossip about

celebrities in the music world. The agent can inform herself and update the acquired knowledge by observing

the web. Fans of musicians can ask for gossip information about stars, bands or people and groups related

to them. This agent is built on top of information extraction, web mining, question answering and dialogue

system technologies. The minimally supervised machine learning method for relation extraction gives the

agent the capability to learn and update knowledge constantly from the web. The extracted relations are

structured and linked with each other. Data mining is applied to the learned data to induce the social network

among the artists and related people. The knowledge-intensive question answering technology enhanced by

domain-speciﬁc inference and active memory allows the agent to have vivid and interactive conversations

with users by utilizing natural language processing. Users can freely formulate their questions within the

gossip data domain and access the answers in different ways: textual response, graph-based visualization of

the related concepts and speech output.

1 INTRODUCTION

The development of information extraction and ques-

tion answering in recent years opens new perspec-

tives for simple but effective interactive dialogue sys-

tems (J¨onsson et al., 2004; Strzalkowski et al., 2005;

Theune et al., 2007). Information extraction enables

dialogue systems to access and understand natural

language texts stored in semi- or unstructured for-

mats, thus, allowing them to make use of the contents

provided by the web, the world’s largest informa-

tion repository. Question answering technology gives

a conversational agent the capability of understand-

ing natural language questions and retrieving answers

from a large knowledge or content pool. At the same

time, question answering systems enhanced by some

dialogue competence enable natural communication

with the human users. The combination of informa-

tion extraction, question answering and dialogue is a

new approach to a conversational agent who is able

to understand natural language questions and provide

answers by extracting and mining information from

a large amount of textual data in structured, semi- or

unstructured form.

One of the hardest challenges in our information

world is to constantly keep the information up to date

and to prepare it in such a way that users can easily

understand and exploit it. We have developed a new

architecture for conversational agent systems that can

learn, update and interpret information from the web

and make conversations with end users, provide an-

swers to their questions and even help them to gain

insights into the application domain. We selected gos-

sip about celebrities in the music world as the domain

for our experimental setup, because many of them ex-

hibit interesting and dynamic aspects with respect to

both their private and professional life. Furthermore,

they are connected to each other in a variety of ways.

Internet news and blogs report on them from differ-

ent perspectives. Our task is to model this domain by

covering relevant facts and trivia on the musicians and

their communities and by discovering new properties

and relations. The acquired information will be uti-

lized as a knowledge resource for conversations with

end users. Users can raise natural language questions

about a special artist or ask for relationships between

115

Xu F., Adolphs P., Uszkoreit H., Cheng X. and Li H. (2009).

GOSSIP GALORE - A Conversational Web Agent for Collecting and Sharing Pop Trivia.

In Proceedings of the International Conference on Agents and Artiﬁcial Intelligence, pages 115-122

DOI: 10.5220/0001663901150122

 SciTePress

artists. Our system provides answers from its knowl-

edge base or even hints at newly discovered informa-

tion.

In comparison to existing systems, our conversa-

tional agent, called “Gossip Galore”, is an active self-

learning system. It starts with only a very small num-

ber of artists and bands and then gradually ﬁnds many

more artists and bands. This is realized by the appli-

cation of a minimally supervised relation extraction

system (see section 3). Users can actively give com-

ments on the answers provided by the agent, which

is useful for self validation. Thus, “Gossip Galore”

contains two major parts: one is the knowledge ac-

quisition component and the other one is the compo-

nent for communication and conversation. Both parts

interact with each other and contribute to the self-

learning process.

The paper is structured as follows: Section 2 de-

scribes the project and the general context in which

“Gossip Galore” is embedded. Section 3 explains the

web mining techniques for the knowledge acquisition,

while section 4 presents the dialogue modelling and

question answering component. Section 5 gives an

overview of the related work. The conclusions and

future steps are described in section 6.

2 RASCALLI

The research presented here is conducted within

the project Responsive Artiﬁcial Situated Cognitive

Agents Living and Learning on the Internet (RAS-

CALLI). RASCALLI is supported by the Sixth

Framework Programme of the European Commission

in the area of Cognitive Systems (IST-27596-2004).

Its goal is to develop and implement cognitively en-

hanced artiﬁcial companions by combining natural

language processing, question answering, web-based

information extraction, semantic web technology and

interaction-driven proﬁling with cognitive modelling

(Krenn, 2008). This work is further supported by the

project KomParse,which is devotedto equipping non-

player characters in computer games with dialogue

capacities.

In the realized system, the RASCALLI agents as-

sist users in extracting information from the web and

other resources. Users can own their own RASCALLI

agents, which are 3D modelled virtually embodied

conversational agents. The perception and action

components of the RASCALLI agents are modelled

by a combination of information extraction, ques-

tion answering and dialogue capabilities. Within the

project some major strands of research are devoted to

the investigation and modelling of architectures that

combine all major components of cognitive systems.

This is an ambitious and demanding task, and as a

step on the way, the results reported here are a prag-

matic compromise that combines state-of-the-art and

novel methods from information extraction, question

answering, semantic technologies and visual anima-

tion with insights from cognitive modelling into a ro-

bust fun application.

3 WEB MINING FOR

KNOWLEDGE ACQUISITION

One of the major competences of the RASCALLI

agents is that they can learn and acquire knowledge

constantly from the web according to user interests.

The minimally supervised machine learning methods

for relation extraction provided by the system DARE

can be easily utilized for realizing this competence

(Xu et al., 2007; Xu et al., 2008a). DARE can be

initialized with several examples of relations about

artists or bands as seed provided by the users and then

learn rules which map the linguistic structures to these

semantic relations. The rules can be applied to texts to

discover new relation instances, which can be reused

as seed again for new rule discovery.

The experimental domain selected for RAS-

CALLI is gossip about celebrities in the pop world.

We start with domain modelling to deﬁne the poten-

tially relevant concepts and relations that will serve

as a framework for the musician proﬁles and the as-

sociated gossip information to be acquired. Given the

relevant concepts and their relations, we apply DARE

to acquire instances of the relations from the web.

3.1 Domain Modelling

The aim of the domain modelling is to identify and

structure the relevant concepts and relations within

the gossip domain. The current domain contains

properties of a musician such as personal proﬁles,

social contexts, achievements, gossip topics and ca-

reer relevant issues. The gossip content is modeled as

an ontology, utilizing the formal language OWL. The

concepts and properties centering on the musician are

depicted in ﬁgure 1.

3.2 Knowledge Acquisition

Many resources on the web report on celebrities, e.g.,

online news sites, Wikipedia, music portals, fan blogs

and forums. The information mentioned above is

stored in different formats: unstructured (free text),

semi-structured (e.g., Wikipedia) or almost structured

ICAART 2009 - International Conference on Agents and Artificial Intelligence

116

Figure 1: Domain ontology (simpliﬁed).

Figure 2: Webmining workﬂow.

(e.g. NNDB). Therefore, we propose a hybrid infor-

mation discovery strategy to detect as much informa-

tion as possible, as shown in ﬁgure 2.

We apply information wrapping, information ex-

traction and information merging techniques to ac-

quire new knowledge. The whole discovery is em-

bedded in a bootstrapping framework, namely, start-

ing with some examples and then learning more and

more information after several iterations.

Relation Extraction with DARE. DARE provides

a general framework for the extraction of relations

and events with various complexities (Xu et al., 2007;

Xu et al., 2008a). This method is minimally super-

vised since the system works with a collection of free

natural language texts without any annotation of do-

main information. The only domain knowledge for

the whole process is the seed. DARE can use lin-

guistic knowledge as it is provided, for example, by

named entity recognizers and linguistic parsers. The

complexity of the seed determines the complexity of

the extracted relations. The seed helps us to identify

the explicit linguistic expressions containing men-

tions of n-ary relation instances or instances of their

k-ary projections where 1 ≤ k < n. Therefore, DARE

can be easily adapted to user interests. Users provide

only some new examples of the relations they are in-

terested in; DARE can learn additional information

from the web based on these examples.

In the current system, we apply SProUT

(Drozdzynski et al., 2004) for the recognition of per-

son names and other concepts (e.g., band and group

names, date time, nationalities, instruments, religions,

sexual orientations) and utilize the Stanford Parser

(Klein and Manning, 2003) to detect linguistic depen-

dency structures. DARE was originally used to ex-

tract information about Nobel Prize winners from free

text. Later experiments showed how to adapt learned

DARE rules for the Nobel Prize award domain to dis-

cover awards won by musicians (Xu et al., 2008b).

Let us look at the following example. Given a seed

example about a Grammy award won by Madonna for

a speciﬁc category in the year 1992:

(1) hMadonna, Grammy, Best Long Form Music

Video, 1992i

The natural language sentence which matches this

seed is:

(2) Madonna won her ﬁrst Grammy in 1992 in the

Best Long Form Music Video category for the

laserdisc release of her 1990 Blond Ambition

Tour.

DARE can extract a linguistic pattern from the

seed example and the matched sentence where the lin-

guistic arguments are associated with their semantic

roles in the semantic relation, after applying linguis-

tic analysis to the sentence. The simpliﬁed DARE rule

looks as follows:

(3) hsubject: recipienti win hobject: prizei hmod:

yeari hmod: categoryi

Information Wrapping. Information wrapping is

responsible for collecting structured data from struc-

tured or semi-structured web sites. It discovers the

HTML structures which indicate the relations deﬁned

in our ontology. We apply this technology to web sites

such as Wikipedia and the special web portal for peo-

ple and their proﬁles, namely, the NNDB.

GOSSIP GALORE - A Conversational Web Agent for Collecting and Sharing Pop Trivia

117

Figure 3: Social network of Madonna.

The method starts with a set of musicians and their

relation instances as seed. Our system sends a query

containing an instance from the seed set as a query to

the web sites and discovers the rules which map the

HTML structures to the relation structures.

Induction of the Social Network. Given the dis-

covered relations among the musicians themselves

and other people, we developed a special system

which can construct a social network from the relation

instances. For example, ﬁgure 3 shows the social net-

work of Madonna. The social network also serves as

the basis for the active dialogue memory of the agent.

4 CONVERSATIONAL AGENTS

In RASCALLI, the central method for users to ac-

cess the acquired knowledge is to communicate with

the user’s personal embodied conversational agent

(ECA). The core functionality of the agent is ques-

tion answering, wrapped in a smooth natural language

dialogue. One main design criterion is to create and

enhance an immersive effect on the user when inter-

acting with the system. The agent should be physi-

cally embodied, she should be situated in a consistent

physical environment, and she should act naturally.

The interaction between users and RASCALLI

can be described as follows. After logging in to

the platform, a three-dimensional visualization of the

user’s agent is displayed (see ﬁgure 4). Just as in an

instant messaging program, the user communicates

with the agent by typing messages into a text ﬁeld.

The agent, on the other hand, responds with natu-

ral language utterances which are presented in their

spoken form (by the use of the open source speech

Figure 4: The Gossip Galore conversational agent.

synthesis system OpenMary (Schr¨oder and Hunecke,

2007)), along with their written form. But the agent’s

means of communication are not restricted to verbal

actions. Complex answers such as the social network

of a star can be visualized on a TV screen, which is

embedded into the agent’s environment. Where it is

appropriate, the agent also emphasizes her responses

by facial and body gestures such as shrugging the

shoulders, nodding or shaking her head and pointing

to the screen.

4.1 Architecture

The RASCALLI system is realized as a server-client

architecture. Users are connected to the server via the

3D client, which displays our ECA and manages the

interaction with the user. The actual control logic of

the agent is executed on the server. The server’s func-

tion is to accept new connections, to manage users

and their logins, and to route messages between the

3D client and the conversational agent.

Figure 5 shows the component hierarchy of our

conversational agents. Some of the components are

responsible for processing various linguistic aspects

of the dialogue, whereas others are concerned with

knowledge representation, management and retrieval

as well as behavioural procedures. Details on the in-

terplay between these components when processing

dialogue turns are presented in the following subsec-

tion.

4.2 Dialogue Processing

When the conversational agent receives an utterance

from the user, its task is to compute a suitable dia-

logue turn in response. We follow a pipeline archi-

tecture for realizing this: the user’s input string is

ﬁrst linguistically analyzed, then it is interpreted in

ICAART 2009 - International Conference on Agents and Artificial Intelligence

118

Figure 5: Components of a conversational agent.

Figure 6: Processing dialogue turns.

the current dialogue context and turned into a suitable

plan for a response action that is executed in the third

stage, leading to an abstract representation for the an-

swer, which is realized with verbal and non-verbal

means in the fourth and ﬁnal stage. This basic data

ﬂow when processing dialogue turns as a response is

depicted in ﬁgure 6. In the following, the four main

components are presented in greater detail.

The idea of having two separate components for

input processing, namely, input analyzer and input in-

terpreter, one for the analysis and one for the interpre-

tation of the user’s input, serves the purpose of draw-

ing a clear boundary between the general and reusable

and the domain-speciﬁc parts of the system. The in-

put analysis component relies on standard domain-

independent linguistic tools, namely a spell checker,

a named entitity recognizer, and a parser producing a

linguistic analysis of the input, for which we currently

employ a fuzzy paraphrase matcher to approximate

the output of a deeper syntactic/semantic parser.

Each utterance is associated with a meaning rep-

resentation as well as with the answer focus and the

expected answer type in case of questions. Note

that by mapping utterances of quite different sentence

types such as plain questions (“Who is Madonna?”),

statements with embedded questions (“I wonder who

Madonna is.”), statements about the user’s inter-

ests without embedded questions (“I’m interested in

Madonna.”) to the same semantic representation, we

can conﬂate sets of user utterances with the same in-

tended meaning.

In the second stage, the input is interpreted in the

current dialogue context, considering previouslymen-

tioned entities for resolving anaphora as well as the

current dialogue state for modelling the system’s ex-

pectations about the user’s turn. If, for instance, a

substring can only be resolved as a named entity with

the help of the spell checker, the system poses a clari-

ﬁcation question to the user and sets the dialogue state

accordingly. This allows the system to interpret a fol-

lowing utterance by the user such as “yes” or “no”,

which would otherwise not be understood. The re-

sult of the input interpretation stage is an abstract plan

to perform a certain action. For example, factoid in-

domain questions result in a plan to look up the data

in our knowledge base, general information requests

about an in-domain person result in the plan to show

a proﬁle page of that person, out-of-domain questions

about a known person result in the plan to present a

suitable web link, and so on.

In the third stage, the Response Handler compo-

nent executes the planned action. For factoid ques-

tions, this means that the corresponding query is

looked up and submitted to the knowledge base. If the

user asks for general information about a person, the

URL of the corresponding proﬁle page is constructed.

The user may also have asked whether there are any

new information about a musician he is interested in.

In this case, an online search for new connections be-

tween people is performed.

The agent follows certain pragmatic principles of

relevance when giving answers to questions. By as-

signing the same semantics to indirect speech acts

(“I wonder who the boyfriends of Madonna are.”) as

to the corresponding direct speech act (“Who are the

boyfriends of Madonna?”), we are able to return a rel-

GOSSIP GALORE - A Conversational Web Agent for Collecting and Sharing Pop Trivia

119

evant response to the user’s request. Similarly to the

previous example, certain yes/no-questions (“Does

Madonna have any boyfriends?”) can be answered

as if they were wh-questions. Instead of giving a sim-

ple “yes” answer, the agent also lists the values for the

queried variable.

Not all of the performed actions necessarily lead

to a satisfying result, though. If no positive answer

can be found for a question or if the question lies out-

side the covered domain, we still want to be able to

provide a constructive answer. If, for instance, the

user asks about a person we do not have detailed infor-

mation about but for whom a Wikipedia entry exists

(e.g. “Tell me something about Nicolas Sarkozy!”),

we point the user to this page, using the embedded

TV screen for displaying the page. If, on the other

hand, the user asks an in-domain question for which

the system does not have any results, we direct the

user to a Google search page with appropriate query

parameters in order to help him ﬁnd relevant informa-

tion.

The outcome of the performed action is always an

abstract representation of the agent’s response. This

might be a simple boolean value for yes/no-questions,

a list of entities for factoid questions, a URL for

the system’s own information services or to exter-

nal sources, and so on. This information is ﬁnally

realized as a communicative act in the fourth stage,

the multi-modal generation. We currently employ

template-based generators for both producing the nat-

ural language utterance as well as for the multi-modal

message with gestures and TV screen commands.

When generating natural language answers to ques-

tions, care has to be taken how these answers are pro-

vided. Since we have to expect that our knowledge

base is incomplete and that the acquired information

could partly be inaccurate (particularly in the gossip

domain), special relativizing expressions such as “ac-

cording to my sources” are produced as part of the an-

swer. The introduction of pronouns for entities men-

tioned before helps making the utterance less static

and the conversation more natural.

4.3 Multimodal Communication

Gossip Galore uses several modalities for communi-

cating with the user. First of all, all the agent’s ut-

terances are spoken, with the help of a speech syn-

thesis system. The verbal part of the answer is addi-

tionally supported by gestures. To make maximal use

of the available means for communication, we also

use a TV screen embedded in the agent’s environ-

ment, which is able to display arbitrary web sites, to

present illustrations of the current answer (see ﬁgure

Figure 7: Multimodal answers – illustrating locations.

Figure 8: Social network visualization.

7) or even to provide the very content of the answer

where the answer is not a single fact or a small set of

facts but would require a complex explanation involv-

ing heterogenous kinds of information (see ﬁgure 4)

or would lead to a rather longish and tiring answer if

it were realized verbally (see ﬁgure 8).

5 RELATED WORK

Web-based question answering systems typically pro-

ceed in several stages: i) the question is turned into a

query for a standard search engine, ii) a set of rele-

vant web sites is retrieved, iii) text passages are se-

lected as an answer from that document set (Neu-

mann, 2008). A variation of this idea is applied in

HITIQA (Strzalkowski et al., 2005), an interactive

open-domain question answering system for complex

exploratory questions, where the answer is retrieved

based on complete semantic event frames which are

matched against the frame of the question in the last of

ICAART 2009 - International Conference on Agents and Artificial Intelligence

120

the three steps above. Thus, much like in RASCALLI,

the system performs a more structured semantic anal-

ysis of the original data (with respect to the question

at hand). In contrast to our system, and more in ac-

cordance with the common web-based QA approach,

however,the extracted information is only used for se-

lecting the most relevant text passages. Furthermore,

it is only used for the current QA task, not for building

up world knowledge.

The approach followed in our system is to learn

new information from a large unstructured text pool

and store it in a knowledge base for structured access.

Such an approach is also followed in BIRDQUEST

(J¨onsson et al., 2004), a QA system for answering

ornithological questions. As in our system, the in-

formation is extracted from natural language text (al-

though from a single source – a bird encyclopedia

– with much stricter conventions than arbitrary web

documents) and then stored in a relational database.

Unlike our system, however, the system is not self-

learning; suitable information extraction patterns are

not learned automatically but have to be provided as

a resource.

The potential beneﬁts and sub-tasks involved in

enhancing question answering with dialogue capaci-

ties to get interactive question answering have been

brieﬂy discussed in the Q&A Roadmap (Burger et al.,

2000). Recourse to a discourse memory for tracking

entities over several questions has played a role in the

context task of the QA track at the TREC 2001 confer-

ence and when processing question series in the main

tasks in the QA tracks of TREC 2004 and 2005. The

extension of question answering to more interactive

dialogue has been tackled in the Complex Interactive

QA (ciQA) task in the QA tracks of TREC 2006 and

2007

There are several projects that enhance a QA

system with more interactive capabilities, namely

BIRDQUEST (J¨onsson et al., 2004), HITIQA (Strza-

lkowski et al., 2005), RITEL (Rosset et al., 2006),

the IMIX demonstrator (Theune et al., 2007), and

SMARTWEB (Reithinger et al., 2007). RASCALLI

has a different focus compared to all of these systems

in that it i) aims to create a personal relationship with

the user by the use of user-adaptive knowledge acqui-

sition methods, and ii) conducts a vivid conversation

with the user that mimics human-to-humancommuni-

cation, creating the immersive effect of a living entity

with its own personality.

Please refer to the TREC homepage at

http://trec.nist.gov/

for further information and

references.

6 CONCLUSIONS

We have described the overall architecture and main

components of a new class of web-based virtual

agents. Although the design of the agents is strongly

inﬂuenced by empirical observations and theoretical

models of natural cognitive agents, our goal has not

been a simulation of biological cognition. This aim is

partially targeted by other research strands within the

RASCALLI consortium. The objective of the demon-

strated architecture and implementation has been a

rather pragmatic and simpliﬁed agent model that ex-

hibits the desired performance properties and serves

as the starting point for a range of extensions and

additional applications. The achieved relevant per-

formance properties are: robustness, accuracy, self-

improvement and nearly real-time behavior.

The planned future extensions include the integra-

tion of deeper language processing methods instead

of or in addition to the fuzzy paraphrase matcher. A

prime candidate for this extension is our own deep

syntactic/semantic parser. Another plan concerns the

required temporal aspects of relations. It is the dy-

namics of the domain that provide the basis for the

gossip. Properties and relationships change quite of-

ten. By detecting and relating the utterance and re-

port times of the various information sources, a mul-

titude of answers may be temporally sorted. Once in

a while, contradicting information is harvested. In

some cases, these contradictions result from an un-

resolved temporal succession, i.e. the contradicting

facts were true at different times. In other cases, one

of the contradicting facts is simply false. In order to

deal with such situations, we need to enrich the infor-

mation extraction by methods for credibility check-

ing, which will be adopted from IE/IR research.

Finally, we plan to exploit the dialogue mem-

ory for moving more of the dialogue initiative to the

agent. In cases of missing or negative answers or in

cases of pauses on the user side, the agent can use the

active parts of the dialogue memory to propose ad-

ditional relevant information or to guide the user to

fruitful requests within the range of user’s interests.

However, the hardest test for the agent architecture

will be the extension to other domains and tasks that

may be less error forgiving than the colorful world of

pop trivia.

ACKNOWLEDGEMENTS

The work presented here was supported by the in-

ternational project RASCALLI funded by the Sixth

Framework Programme of the European Commission

GOSSIP GALORE - A Conversational Web Agent for Collecting and Sharing Pop Trivia

121

in the area of Cognitive Systems (IST-27596-2004),

and partially funded through a grant to the project

KomParse by the ProFIT programme of the Federal

State of Berlin and the EFRE programme of the Eu-

ropean Union. We are also grateful to the coopera-

tion with the HyLap project funded by the German

Ministry for Education and Research (BMBF, FKZ:

01 IW F02). Many thanks go to our RASCALLI

project partners, in particular, Radon Labs team, led

by Nicolaas Bongaerts, for the development of the 3D

client, and Brigitte Krenn and her team from OFAI

and SAT as well as Rebecca Dridan from the Depart-

ment of Computational Linguistics at the University

Saarbr¨ucken for their suggestions and comments.

REFERENCES

Burger, J., Cardie, C., Chaudhri, V., Gaizauskas, R.,

Harabagiu, S., Israel, D., Jacquemin, C., Lin, C.-Y.,

Maiorano, S., Miller, G., Moldovan, D., Ogden, B.,

Prager, J., Riloff, E., Singhal, A., Shrihari, R., Strza-

lkowski, T., Voorhees, E., and Weishedel, R. (2000).

Issues, tasks and program structures to roadmap re-

search in Question & Answering (Q&A).

Drozdzynski, W., Krieger, H.-U., Piskorski, J., Sch¨afer, U.,

and Xu, F. (2004). Shallow processing with uniﬁca-

tion and typed feature structures – foundations and ap-

plications. K¨unstliche Intelligenz, 1:17–23.

J¨onsson, A., And´en, F., Degerstedt, L., Flycht-Eriksson, A.,

Merkel, M., and Norberg, S. (2004). Experiences

from combining dialogue system development with

information extraction techniques. In Maybury, M. T.,

editor, New Directions in Question Answering, pages

153–168. MIT Press.

Klein, D. and Manning, C. D. (2003). Accurate unlexical-

ized parsing. In Proceedings of the 41st Meeting of

the Association for Computational Linguistics (ACL

2003), pages 423–43.

Krenn (2008). Responsive artiﬁcial situated cognitive

agents living and learning on the internet. Poster pre-

sented at the International Conference on Cognitive

Systems (CogSys 2008).

Neumann, G. (2008). Strategien zur Webbasierten Multi-

lingualen Fragebeantwortung: Wie Suchmaschinen zu

Antwortmaschinen werden. Computer Science - Re-

search and Development, 22(2):71–84.

Reithinger, N., Herzog, G., and Blocher, A. (2007).

SmartWeb – mobile broadband access to the seman-

tic web. KI – K¨unstliche Intelligenz, 2/2007.

Rosset, S., Galibert, O., Illouz, G., and Aur´elien, M. (2006).

Integrating spoken dialog and question answering: the

Ritel project. In Proceedings of INTERSPEECH 2006.

Schr¨oder, M. and Hunecke, A. (2007). Mary tts participa-

tion in the Blizzard Challenge 2007. In Proceedings

of the Blizzard Challenge 2007, Bonn, Germany.

Strzalkowski, T., Small, S., Hardy, H., Yamrom, B., Liu,

T., Kantor, P., Ng, K., and Wacholder, N. (2005). HI-

TIQA: A question answering analytical tool. In Pro-

ceedings of the International Conference On Intelli-

gence Analysis (IA-05), McLean, VA.

Theune, M., Krahmer, E., van Schooten, B., op den Akker,

R., van Hooijdonk, C., Marsi, E., Bosma, W., Hofs,

D., and Nijholt, A. (2007). Questions, pictures, an-

swers: Introducing pictures in question-answering

systems. In Tenth international symposium on social

communication, pages 450–463, Cuba.

Xu, F., Uszkoreit, H., and Li, H. (2007). A seed-driven

bottom-up machine learning framework for extracting

relations of various complexity. Proceedings of the

45th Annual Meeting of the Association of Computa-

tional Linguistics, pages 584–591.

Xu, F., Uszkoreit, H., and Li, H. (2008a). Task driven

coreference resolution for relation extraction. In Pro-

ceedings of the European Conference for Artiﬁcial In-

teligence ECAI 2008, Patras, Greece.

Xu, F., Uszkoreit, H., Li, H., and Felger, N. (2008b). Adap-

tation of relation extraction rules to new domains. In

Proceedings of the Sixth International Conference on

Language Resources and Evaluation (LREC 2008).

ICAART 2009 - International Conference on Agents and Artificial Intelligence

122