Tutorial
Presented at User Modelling 2007:
Affective Natural
Language Generation
The handouts for this
tutorial are available at http://www.csd.abdn.ac.uk/~cmellish/um2007/handouts-final.pdf
1. Tutorial Chairs
Fiorella de Rosis
Dept of Informatics
University of Bari
70126 Bari, Italy
Tel: +39 080 5443282
Email: derosis@di.uniba.it
Chris Mellish,
Computing Science,
University of Aberdeen,
King’s College,
Aberdeen AB24 3UE, UK
Tel: +44 (0)1224 272293
Email: cmellish@csd.abdn.ac.uk
2.
Topic, Goals, Relevance and Significance to UM
Natural language generation (NLG) is the task of automatically producing appropriate human language based on non-linguistic inputs. Affective NLG is where NLG meets affective computing, to produce language that achieves emotional or other non-rational effects on the reader. In fact, all text produced by a computer has non-rational effects on its readers, though these are usually unintentional, because of limitations in our existing models of linguistic communication. So really affective NLG is about how we can become more aware of those effects that language has and intentionally manipulate them.
The goal of this tutorial is to convey the potential of
affective NLG, describe some of the initial work done in this area and to
outline some of the challenges it makes to user modelling. Individual users
differ in how they respond emotionally to a given piece of natural language,
and so one of the challenges for user modelling is to be able to capture this
variation in order to support choices by computer systems between linguistic
alternatives in their utterances.
3.
Target Audience
The tutorial is relevant to workers in HCI and other areas who are concerned with personalised adaptive interfaces which generate language of some kind (whatever the complexity of their generation models). The tutorial will be introductory and will presuppose only limited previous knowledge of NLG or affective computing. It will therefore be suitable to people from a range of disciplines, including Artificial Intelligence, HCI and Education.
There is a significant overlap between NLG researchers and user modellers (for instance, in UM’07, at least one of the program co-chairs and 9 of the programme committee have worked in NLG), and so there is evidence that a topic of this kind, which spans the two fields and requires new and coordinated developments in both, will be of interest to UM attendees. Considering the increasing interest of UM participants in the two areas that this Tutorial combines (affective computing and NLG), we expect to have a good number of participants.
4. Organisation of the Tutorial
4.1 Format
The Tutorial will not present just the union of work in affective
computing and NLG: it will be the result of a coordinated synthesis developed by
the two organisers. It will combine a state of the art description with a
vision of the possible trends of this fast evolving domain. Description of
methods will be integrated with critical reflection on the main results
obtained, of their limits and of the open problems, so as to solicit active
contributions from the participants.
A short questionnaire will be administered to those who register for
the Tutorial, to understand their interests and backgrounds, so that the
lecturers can adapt their content accordingly. This will enable us to remain
within the limited time available and yet go deeper into the topics which are of
interest for the participants.
4.2 Content will include some of
the following
Affective NLG: The
state of the art
·
Definitions
of affect-related concepts; scope and goal of affective computing.
·
Affective
NLG as (1) portraying a writer with a given emotion or (2) inducing an emotional
effect on the hearer. Possible applications: advice-giving systems, intelligent tutoring, call centers
and others. Some examples: STOP; why STOP might have failed.
·
Overall
state of the art in affective user modeling and with NLG generally and NLG architectures/
terminology.
·
Summary
of knowledge about language use and affect (e.g. how language reveals deeper
things about the speaker: Pennebaker et al)
- General ideas of how to produce affective NL messages: in content, style, dialog dynamics.
- General method: copy expert human writers, though without necessarily understanding why they do it or what they are responding to, and/or mimic the reasoning experts do about the addressee’s mental processes.
- Description of specific work, e.g. conveying doubt (Carberry), personality (Isard), empathy and social attitude (Hovy, , …) through NLG;
Research
challenges for affective NLG and user modelling
- Modelling stable and unstable affective characteristics of the user (e.g. Belief Networks and their limitations).
- Reasoning about what effect to achieve (planning, integration with BDI models).
- Handling individual variation in the use of words and response to words.
- Tackling acquisition (implicit vs explicit), consistency, validation and privacy problems in affective user models.
- Choosing a strategy for inducing an effect: the ability to portray a writer with an emotion gives a basis for exploiting empathy, reciprocation, contrast; other strategies involve appropriate choice of content (e.g. for persuasion).
- Evaluating success.