|
Veranstalter
|
|
Christopher Habel,
Carola Eschenbach,
Wolfgang Menzel
|
|
Zeit/Ort
|
|
Di 16-18 F-534 |
|
Inhalt
|
|
Vorträge über Vorhaben und Ergebnisse von Bachelor-, Master- und Diplomarbeiten,
laufenden Dissertationen und Drittmittelprojekten sowie von anderen Forschungsarbeiten aus
dem Bereich der Wissens- und Sprachverarbeitung. Insbesondere wird der interdisziplinäre
Charakter des Forschungsschwerpunktes berücksichtigt, d.h. die Integration von Ansätzen der
Informatik, Linguistik, Logik und Psychologie steht im Vordergrund der Arbeiten. |
|
Termine
|
|
| 13.04.2010 |
Kris Lohmann und Matthias Kerzel
Generating Verbal Assistance in Tactile-Map Explorations
Tactile maps offer access to spatial-analog information for visually
impaired people. In contrast to visual maps, a tactile map has a lower
resolution and can only be inspected in a sequential way, complicating the
extraction of spatial relations among distant map entities. Verbal assistance
can help to overcome these difficulties by substituting textual labels with
verbal descriptions and offering propositional knowledge about spatial
relations. Like visual maps, tactile maps are based on visual, spatial-analog
representations that need to be reasoned about in order to generate verbal
assistance. In this talk, we present an approach towards verbally assisting
virtual-environment tactile map (VAVETaM) realized on a computer system
utilizing a haptic force-feedback device. We present our current research on
understanding the user's Map Exploratory Procedures (MEPs), exploiting the
spatial-analog map to anticipate the user's informational needs, reasoning about
optimal assistance by taking assumed prior knowledge of the user into account
and generating appropriate verbal instructions and descriptions to augment the
map.
|
|
|
| 20.04.2010 |
Jörg Didakowski (Berlin-Brandenburgische Akademie der Wissenschaften)
Finite-State Weighted Constraint Dependency Grammar - Syntactic Dependency Parsing in Linear Time
In the parsing system Weighted Constraint Dependency Grammar (WCDG)
syntactic dependency parsing is formulated as Constraint Optimization Problem
(COP) where well-formedness rules are written as defeasible constraints enabling
partial parsing, structural preferences and degrees of grammaticality. Many
algorithms are explored in order to solve such COPs efficiently. Despite
substantial improvements e.g. by using transformation based techniques the
running time of parsing remains a big problem. In my presentation Finite-State
Machines (FSM) are proposed to represent and solve such COPs within an extended
finite-state approach in order to tackle this drawback. Solving a COP is known
to be NP-hard in general. To avoid such a bad behavior a problem decomposition
technique is worked out where a COP is decomposed tree-like and solved bottom-up
and where different potential decompositions of the problem can be handled in
parallel. Together with a bounded depth of the decomposition's tree structure
the parsing can actually be performed in linear time and space. I will show that
solving COPs by means of FSMs is a promising alternative to improve parsing
run-time.
|
|
|
| 27.04.2010 |
Christine Upadek
Ähnlichkeitsvergleich von Texten mithilfe der Linkstruktur der Wikipedia
Verfahren, die inhaltliche Ähnlichkeit von Texten berechnen, können für
verschiedene Aufgaben eingesetzt werden. Dazu gehören unter anderem
Textkategorisierung, Paraphrasenerkennung und das Vorschlagen ähnlicher Texte zu
einem gegebenen.
Dieser Vortrag gibt einen Überblick über verschiedene
Verfahren, die Textähnlichkeit berechnen. Vorgestellt werden Verfahren aus dem
Information Retrieval und wissensbasierte Verfahren, die Taxonomien aus
lexikalischen Ressourcen wie WordNet oder GermaNet nutzen. Zudem werden
verschiedene Verfahren vorgestellt, die für den Ähnlichkeitsvergleich die
Wikipedia verwenden. Hierfür wird auf den Aufbau der deutschen Wikipedia
eingegangen und am Beispiel des Programms "Wikipedia
Preprocessor" gezeigt, wie die benötigten Informationen aus der
deutschen Wikipedia extrahiert werden können. Außerdem werden ein eigenes
Verfahren, das Artikel, Kategorien und Links zwischen Seiten der Wikipedia
benutzt, um Textähnlichkeit zu berechnen, und die Ergebnisse der durchgeführten
Evaluation vorgestellt.
|
|
|
| 04.05.2010 |
Stefan Zimmermann
Wissensmanagement mit Wikis - Optimierungsprozesse
Wissensmanagement ist in den letzten Jahren in aller Munde. Ein neuer
struktureller Wandel führt die Unternehmenswelt hin zu einer Informations- und
Wissensgesellschaft. Die Erzeugung von Wissen, seine Verbreitung, Nutzung und
Wiederverwendung bekommt immer mehr Gewicht, da sie für die Aufrechterhaltung
der Wettbewerbsfähigkeit unumgänglich geworden sind. Es stellt sich die Frage,
wie Unternehmen an diesen neuen Aspekt des Wissensmanagements herangehen und ihn
in die bestehenden Geschäftsprozesse eingliedern. Was kann ein Unternehmen tun,
um das vorhandene Wissen der Mitarbeiter zu benutzen? Wodurch sind Mitarbeiter
angeregt, das Wissen zu verteilen? Wie kann es gelingen, dass es optimal genutzt
wird, indem es Verbreitung und Wiederverwendung findet? Wie schafft es ein
Unternehmen, das Wissen zu verwalten und dessen Existenz greifbar zu
machen?
Die vorliegende Diplomarbeit entstand in Kooperation mit einem
mittelständischen Softwareunternehmen. In der Arbeit wird analysiert, wie das
Unternehmen die oben genannten Fragen beantwortet. Der Leser soll einen Einblick
bekommen, welche Möglichkeiten diesbezüglich bestehen und wie das Unternehmen
den Umgang mit Wissen bewältigt. Dazu wurden einzelne Elemente des
Wissensmanagements analysiert, um herauszufinden, wo sich das Unternehmen noch
verbessern kann. Das Social-Networking-Prinzip der Wikis macht im Unternehmen
einen entscheidenden Teil des Wissensmanagement-Prozesses aus. Es wurde geprüft,
inwieweit diese Methode im Wissensmanagement Anwendung finden kann und ob es
noch Verbesserungspotential gibt. Dazu sind konkrete Maßnahmen vorgestellt und
miteinander verglichen worden.
|
|
|
| 10.05.2010, 17 ct, B-201 |
Sondertermin im Rahmen des Fachbereichskolloquiums
Maite Taboada
A lexicon-based approach to sentiment analysis
Sentiment analysis is the automatic extraction of information about
opinion and subjectivity from text and speech. In this talk, I describe our
current research in sentiment analysis. The Semantic Orientation CALculator
(SO-CAL) uses dictionaries of words annotated with their semantic orientation
(polarity and strength), and incorporates intensification and negation. I
describe the process of dictionary creation, and our use of Mechanical Turk to
check dictionaries for consistency and reliability. SO-CAL is applied to the
polarity classification task, the process of assigning a positive or negative
label to a text that captures the text's opinion towards its main subject
matter. I show that SO-CAL's performance is consistent across domains
and in completely unseen data. I also describe current research on using
discourse information to improve performance.
|
|
|
| 18.05.2010, F-334 |
Gemeinsamer Vortrag im Rahmen der Oberseminare von TAMS, WSV und NatS
Prof. Xiaoyan Zhu, State Key Lab of Intelligent Technology and Systems, Tsinghua University Beijing, China
Q & A Based Internet Information Acquisition
The Internet, providing the largest databases and encyclopedias of the
information society, has outweighed all other media as a source of information.
Therefore, the way to access the internet to acquire appropriate information has
become more and more important. As a promising information acquisition
technique, question answering is very active in information retrieval and
natural language processing research. This talk introduces recent progress on a
question answering system and its role in internet information acquisition,
addresses the main topics in this area, and focuses on two main issues: text
mining and summarization.
Professor Xiaoyan ZHU got her bachelor degree at
the University of Science and Technology Beijing in 1982, her master degree at
Kobe University in 1987, and her Ph. D. degree at the Nagoya Institute of
Technology, Japan in 1990. She has been teaching at Tsinghua University since
1993. She is deputy Head of the state key lab of intelligent technology and
systems, director of the Tsinghua- HP Joint research center and director of the
Tsinghua-Waterloo Joint research center, Tsinghua University. She is
International Research Chair holder of IDRC, Canada, since 2009. She was deputy
head of the Department of Computer Science and Technology, Tsinghua University
from 2004-2007. Her research interests include intelligent information
processing, machine learning, natural language processing, query and answering
systems and bioinformatics. She has authored more than 100 peer-reviewed
articles at leading international conferences (SIGKDD, ICDM, PAKDD, CIKM, ACL,
APBC) and in journals (Int. J. Medical Informatics, Bioinformatics, BMC
Bioinformatics, Genome Biology and IEEE Trans. on SMC).
|
|
|
| 01.06.2010 |
Isabelle Streicher
Semantic processing of local and directional verbal modifiers in route instructions
The verbal inventory of route instructions mainly consists of position and
motion verbs. Position and motion verbs complement with local and directional
arguments respectively. The syntactic driven, compositional combination of the
verbs' and the arguments' meaning is unambiguous, cause
lexically fixed.
Besides obligatory arguments a verb may combine with
optional modifiers additionally. In contrast to semantic
verb-arguments-combination, semantic verb-modifier-combination can't be
fixed lexically and is therefore less obvious to achieve. In formal semantics, a
standard way of verbal modifier handling is Davidson's approach of
situation modification.
Since Davidson's approach fails in
numerous cases of local and directional modified verbal phrases in route
instructions, I headed for an alternative, systematic, domain specific handling
of the problem during my master thesis. My results will be presented and put up
for discussion in this talk.
|
|
|
| 15.06.2010 |
Niels Beuck
Anticipatory Incremental Parsing in Multi-Modal Context - Partial Analyses and Local Ambiguities
In human-robot and human-computer interaction, natural language
processing systems are confronted with real-time dialog situated in
extra-lingual context. These situations provide very different challenges
compared with the classical processing of digital text like in text mining. On
the one hand, there is no need for a very high throughput, as utterances do not
need to be processed much faster than they are produced by the human
interlocutor, but on the other hand processing needs to start before the end of
production of the utterance to prevent unnecessary pauses. Furthermore,
connections to the context and feedback opportunities like eye movement need to
be evaluated at each point during the processing, to provide a fluid and natural
dialog experience for the human. To meet these requirements, an incremental
dialog system needs to be able to provide partial analyses for partial language
input to other modules like visual processing and action planning, and also to
integrate input from those modules at each point in time.
In this talk I
will present the current state of my dissertation project. The goal of my
project is to design an incremental natural language processing system that
generates meaningful partial analyses and provides an interface for the
interaction with other processing modules. I will give an overview over the
challenges in incremental NLP, especially different strategies to deal with
local ambiguities and how to provide and evaluate partial analyses.
|
|
|
| 22.06.2010 |
Jan Christian Krause
Using Thematic Grids to Document Web Service Operations
Web Services are frequently used for system integration in business
contexts, e.g. workflow automation. Therefore their documentation is required to
be precise and complete. This talk discusses state-of-the-art approaches to
document Web Services and provides examples of the negative impact a lack of
construction guidelines has on the documentation of what the service does. A
verb-focused approach to documentation construction is presented which is based
on the linguistic concept of thematic roles and grids. It is complemented by an
empirical study, showing that the prerequisites of the approach are satisfied.
Finally a concept for a future application of the described approach in the area
of Web Service Orchestration is presented, whose development is the goal of my
dissertation project.
|
|
|
| 29.06.2010 |
Kris Lohmann und Matthias Kerzel
Verbal Assistance in Tactile-Map Explorations
Tactile maps are a means to communicate spatial knowledge providing access
to spatial representations of knowledge for visually impaired people. However,
compared to visual maps, tactile maps have some major drawbacks concerning the
integration of information due to the need of sequential exploration. Verbal
descriptions providing abstract propositional knowledge have an advantageous
effect on tactile map reading. They can be used to communicate knowledge that on
a visual map is usually realized in the form of textual labels. Further, verbal
assistance can facilitate the acquisition of global spatial knowledge such as
spatial relations of streets and support the tactile-map user by assisting
exploration, for example, by giving information about landmarks next to a
street. We present an approach towards a verbally assisting virtual-environment
tactile map (VAVETaM), which provides a multimodal map, computing situated
verbal assistance by categorizing the user's exploration movements in semantic
categories called MEPs. Three types of verbal assistances are discussed. VAVETaM
is realized using a computer system and the PHANToM® desktop haptic
force-feedback device, which allows haptic exploration of 3D-graphics-like
haptic scenarios.
|
|
|
| 02.07.2010, 14 ct, F-334 |
Sondertermin am Freitag
Torsten Hahmann
Contact algebras for verification and design of mereotopologies
Building and verifying mereotopological theories which capture
mereological (parthood) and topological (connection) relations between regions
of space plays a central role in Qualitative Spatial Reasoning. We treat a large
classe of common mereotopologies (those with unique closures) as contact
algebras defined over bounded lattices equipped with a unary operation of
complementation. This algebraic framework extends previous works twofold: (1) it
includes non-distributive contact algebras, in particular the so-called Stonian
p-ortholattices, and (2) it allows establishing necessary conditions of what
constitutes a contact algebra that admits a spatial interpretation. Moreover,
different ontological choices are directly related to algebraic properties. It
turns out that spatially representable contact algebras fix many of these
properties while only few real choices remain. Thus, the framework allows us to
extract three weakest, possibly spatially representable, mereotopologies.
In the second part of the talk, we give an equational axiomatization of the
(non-distributive) Stonian p-ortholattices and show that the theory - and thus
the original theory - exhibits unintended models. Approaches to extend the
axiomatization in order to eliminate unintended models are presented. Other
benefits of the equational axiomatization are significant speed-ups for some
model construction and theorem proving tasks using standard theorem provers. In
the broader scope, our work demonstrates how mathematical representations of
ontologies can help us fully understand, verify, and design ontologies.
|
|
|
| 06.07.2010, 16:15, F-334 |
Sondertermin: Disputationsvortrag
Patrick McCrae
A Computational Model for the Influence of Cross-Modal Context upon Syntactic Parsing
Ambiguity is an inherent property of natural language. Despite the high
frequency and diversity with which ambiguity occurs in unrestricted natural
language, most ambiguities in inter-human communication pass unnoticed, mainly
because human cognition automatically and unconsciously works to resolve
ambiguity. A central contribution to this automatic and unconscious
disambiguation is the integration of non-linguistic information from cognitively
readily available sources such as world knowledge, discourse context and visual
scene context.
In this talk I present a cognitively motivated computational
model for the cross-modal influence of visual scene context upon natural
language understanding. In line with Jackendoff's Conceptual Semantics, I argue
for a model that employs semantic mediation to establish cross-modal referential
links between words in the linguistic input and asserted entities in a
representation of visual scene context (context model). The proposed framework
assigns cross-modal referential links on the basis of conceptual compatibility
between the concepts activated in the linguistic modality and the concepts
instantiated in the context model. The implementation of the model centres
around WCDG2, a symbolic weighted-constraint dependency parser for German.
Situation-invariant semantic knowledge, including semantic lexical knowledge and
world knowledge, is encoded in an OWL ontology (T-Box). The situation-specific
visual scene information in the context models is represented in terms of
concept instantiations from the ontology joined by thematic relations (A-Box). A
predictor component computes acceptability scores for the assignment of semantic
dependencies in the linguistic input given a representation of visual scene
context.
In addition to the model motivation and specification, I report
experiments that demonstrate the effectiveness of the framework. The
experimental findings show that the model successfully integrates visual context
information to effect syntactic disambiguation in notoriously hard-to-parse
cases such as PP attachment, subject-object ambiguity of German plural nouns and
genitive-dative ambiguity of German feminine nouns.
|
|
|
| 13.07.2010 |
Ogeigha Koroyin
The Instruction of Artificial Agents via Controlled Language
Human Agents give and process instructions in natural languages relatively
effortlessly, but often have difficulties - especially if they are
non-specialists - using formal languages. Although, on the other hand, natural
languages are very expressive and highly flexible, they possess very high
degrees of ambiguity and complexity which restrict automatic reasoning severely.
Therefore, the formulation of route instructions to Artificial Agents, for
example the Geometric Agent, with the aid of a natural language is indeed
intuitive for non-expert users but is a large stumbling block for automatic
processing.
In order to support a notation that is intuitive for users and
simultaneously automatically processable for computers, controlled languages
have been designed for diverse domains. In my thesis, I design a Controlled
Language (CERI: Controlled English for Route Instructions) that may be used to
express route instructions for the Geometric Agent.
|
|
