Information on how to find W107 is available on our contact page.
| 20 January 2005 |
Karina Rodriguez ITRI, Brighton University Knowledge Based System to Support Collaborative Product Development Abstract | |
| 27 January 2005 *** Cancelled *** |
Graham Wilcock Imperial College, London XML-based Natural Language Generation | |
| 3 February 2005 |
No seminar | |
| 10 February 2005 |
Kristiina Jokinen Computer Laboratory, Cambridge University Adaptation and interaction for mobile speech interfaces Abstract | |
| 17 February 2005 |
Jonathan Ginzburg Computer Science, King's College London On some differences between dialogue and multilogue: evidence from non-sentential utterances | |
| 24 February 2005 |
No seminar | |
| 3 March 2005 *** Postponed to 24 March 2005 *** |
Rob Koeling Informatics, Sussex University | |
| *** NB FRIDAY 3PM*** 11 March 2005 |
Richard Breheny Department of Linguistics, University College London Defaults in Pragmatics Abstract | |
| 17 March 2005 |
Michel Généreux ITRI, Brighton University Example-Based Semantic Parsing of Conversational Speech Abstract | |
| 24 March 2005 |
Rob Koeling Informatics, Sussex University Domain-Specific Sense Distributions and Predominant Sense Acquisition Abstract | |
| 14 April 2005 |
Prof Nigel Shadbolt Southampton University Web Intelligence Abstract | |
| 12 May 2005 |
Alex Lascarides Informatics, Edinburgh University Probabilistic Head-Driven Parsing for Discourse Structure Abstract | |
Companies have adopted a geographically distributed working
approach in order to remain competitive in a global market. This global
engineering environment has led to the distribution of product life cycle
knowledge affecting the collaboration throughout product development.
Although information technologies, such as the Internet, provide a partial
solution to support such collaboration, there is still a need to support
decision making by providing the right information and knowledge in the
place, time and format required by the geographically distributed companies.
This research proposes the use of a digitalised representation of product
life cycle knowledge to support the development of injection moulding
products. Such representation is the kernel of a Knowledge driven
Collaborative Product Development (KdCPD) system to support the effective
collaboration in a global engineering environment.
Kristiina Jokinen
Adaptation and interaction for mobile speech interfaces
As the state-of-the-art speech and language technology has reached a level where the users can have short conversations with the system, interactive speech-based applications have become more common, and also commercially viable. Applications deal with directory enquiries, information seeking and booking systems, call centers, etc. However, interactions are usually characterised by simple predefined interaction patterns which often leave the user frustrated due to their repetitive and unhelpful nature.
In this talk, I will discuss possibilities for making interaction
more flexible and natural by taking some human language capabilities
into account. I will discuss various approaches to modelling natural
conversations, and focus especially on adaptive user modelling so as
to take the user's expertise level into account. I will review the
research done in the EU-project DUMAS, on adaptation and interface
design, and provide an overview of the AthosMail system, a speech-
based interactive email-application.
Michel Généreux
Example-Based Semantic Parsing of
Conversational Speech
Abstract- Most research in NLP has concentrated on processing data at a low level, such as morphology, part-of-speech tagging and syntactic parsing. Semantic parsing, defined as the mapping of a natural language sentence to its semantic representation, has been limited to isolated tasks, such as word sense disambiguation and information extraction. Other attempts at semantic parsing often rely heavily on syntactic parsing. Syntactic parsing is time-consuming, and for dialogue systems, semantic processing of the spoken input must be done more efficiently and reliably.
In this talk I will present an empirical method for mapping speech input to shallow semantic representation. The semantic parser is based on examples of sentence-meaning pairs and uses context as a "cheap" substitute for syntax. Semantic parsing is realized through a bottom-up type parsing paradigm where the operators are based on semantic concepts, obtained from a lexicon. A statistically trained model specializes the parser, by guiding the runtime beam-like search of possible parses.
The parser was built in two stages, using two different corpora, and was
progressively equipped to handle discourse phenomena. In the first stage,
the parser uses First-Order Logic annotation and is trained on a simple
task-oriented German corpus for querying and browsing newspapers. In the
second stage, it uses Discourse Representation Structure annotation and is
trained on a more complex domain-oriented English corpus for the life and
works of the Danish fairy-tale writer Hans Christian Andersen.
Rob Koeling
Domain-Specific Sense Distributions and Predominant Sense
Acquisition
Distributions of the senses of words are often highly skewed. This fact can be exploited by word sense disambiguation (WSD) systems to back off to the predominant sense of a word when contextual clues are not strong enough. WSD can also benefit from information about the domain of a document, since for many words their meaning is strongly correlated with the domain of the document they appear in. We would therefore ideally like a large manually annotated corpus in every domain of interest so we could derive domain-specific predominant senses for every word. This is clearly impractical.
In this paper we describe the construction of three sense annotated corpora in different domains for a sample of English words, and show quantitatively that: (1) the sense distributions of the words differ depending on domain, and (2) sense distributions are more skewed in domain-specific text.
We then apply our method for acquiring predominant sense information automatically from raw text (McCarthy et al. 2004), and show that: (1) acquiring such information automatically from a mixed-domain corpus is more accurate than deriving it from SemCor, and (2) acquiring it automatically from text in the same domain as the target domain performs best by a large margin.
These results are potentially important for scaling up and widening
the applicability of WSD systems.
Richard Breheny
Defaults in Pragmatics
There has been a recent revival of a debate as to the status of common
conversational implicatures (sometimes called 'Generalised Conversational
Implicatures'). In particular, Levinson (2000) and Chierchia (2004) have
argued that these inferences are default not only in the sense of being
available in the absence of evidence to the contrary but that they are
computed by default - either through the compositional semantic component of
the grammar or within a pragmatics sub-module. Scepticism for this view had
been expressed by Carston (1998) and others who take an opposite but equally
extreme position as to the status of the implicatures - that they are only
available where fully warranted by the utterance context and that they are
computed as nonce inferences on each occasion. In this talk I will present
the results of some on-line reading-time studies which test the predictions
of these opposing accounts. Our results systematically disconfirm the strong
default view and are consistent with the strong context-driven view. The
same story emerges from other recent studies which use different
methodologies and also from child data. I will then explore the possibility
of a more moderate approach based on constraint-based approaches to sentence
processing (McDonald 1994, Altmann 1998) drawing on our own studies in
lexical interpretation and recent constraint-based research.
Nigel Shadbolt
Web Intelligence
The extraordinary human construct that is the World Wide Web is a truly Disruptive Technology. There are now hundreds of millions of users, billions of indexed web resources, it is used in every country on Earth and yet only a tiny percentage of users is "trained" in any way. This remarkable construct is both massively distributed and largely open.
With this amount of content and usage the integration of information across space and time leads to new opportunities. From on-line shopping to collaborative e-Science the web is changing how information is generated, deployed and used. The military will have few options but to take advantage of the huge investment that the commercial and research sectors will make in web service solutions and architectures.
This lecture will examine the extent to which intelligent web services are
evolving to cope with diverse sources of information on a global scale.
It will examine the particular way in which Artificial Intelligence is
being woven into the web. It will review how these developments are likely
to shape future military capabilities.
Alex Lascarides
Probabilistic Head-Driven Parsing for Discourse Structure
I will present joint research with Jason Baldridge.
We propose a data intensive approach to building interpretable
discourse structures for appointment scheduling dialogues. Our
strategy involves a tree-based representation of discourse over which
we define several probabilistic head-driven parsing models. We
evaluate performance on recovering labelled and unlabelled discourse
relations. Our results show that discourse-based information, such as
turn-taking and domain specific goals, has a large positive impact on
model performance. The best model performs significantly better on
labelled relations than a baseline that connects an utterance to its
context in the most frequent way.