This funding from the Engineering and Physical Sciences Research Council under its "platform grant" scheme provides general support for the activities of the NLG group, focussing on three strands in particular:

  • experimental studies of how readers are affected by language
  • modelling of language users, particularly with regard to "affective" aspects
  • examination of how best to construct more general NLG systems, in terms of internal structures and processes

The grant has supported a number of workshops, collaboration with research work elsewhere, further development of the SimpleNLG software, and several "mini-projects" in the area of NLG.

 

Staff

 

What the work is about

Limitations of Traditional NLG

In the real world, texts vary enormously both in their communicative purpose, and in the abilities and preferences of the people who read them. Much previous research in NLG has assumed that the purpose of generated texts is simply to communicate factual information to a user [17]. There has been little attention to other aims, such as persuading people [16], teaching people [9,25], helping people make decisions [18], [6], and entertaining people [19]. While texts with these other aims usually do communicate information, they do so in order to affect the reader at a deeper level, and this has an impact on how the information should be communicated (the central task of NLG). Even where the main goal is to inform, the other ways in which the language affects the reader may have an important effect on the achievement of that goal.

Traditional NLG tackles a single type of generic goal (factual information) for a general user (or one of a small number of user types). The focus needs to be broadened to a variety of types of goals for specific users. Although NLG research has begun to explore the issues of reader variability (eg [23], [1]), including user modelling (see [24] for a good review), this is at an early stage, and tends to concentrate on broad decisions about content rather than fine-grained linguistic form, the focus of our proposed work.

Our own projects have begun to address these issues. User groups have included children with linguistic difficulties (STANDUP), adults with limited literacy (SkillSum), general members of the public (STOP, ILEX [12]), and professional doctors and engineers (SumTime, [6]), sometimes with individual customisation (STOP, SkillSum). The texts have been informative (SumTime), persuasive (STOP, SkillSum), humorous (STANDUP), and entertaining (NECA).

 

Strategic Vision

NLG has enormous potential to achieve benefits in the real world, especially given the growing importance of eCommerce, eHealth and eGovernment, but current NLG applications exist only in niche areas. We believe that there are two main reasons for this:

  1. Firstly, many real applications challenge the assumptions of traditional NLG highlighted above (single, generic goal; general user). We would like to push forward the scientific understanding of how the attributes of an individual reader (and the reading process for them) influence the effect that particular linguistic choices have on them. This will then result in an ability to build systems which, from a model of the reader, can intelligently select linguistic forms in order to achieve increasingly ambitious effects. Hence our goal is to learn better how to affect people with natural language.
  2. Secondly, NLG can be somewhat inward-looking. As our current projects (PolicyGrid, BabyTalk) show, NLG adds value to other computational solutions and often cannot be viewed as a stand-alone technology. We would like to lead in the emergence of NLG from its small corner, as it contributes to wider research initiatives and is increasingly exploited commercially. This requires us to make use of the methodologies and knowledge of other disciplines, within and outside Computer Science, to a much greater extent than hitherto. Hence there is a need for strategic alliances with a variety of researchers and disciplines.

To address the problems highlighted above, we see the following scientific themes as especially relevant:

  1. Psychology and Reader Experiments.We need to understand the relevance to NLG of attention, perception and memory. Particularly relevant are results about human reading [15] and how humans align their language use in order to effectively reach their hearers [2]. Although we are already at the forefront of measuring the effects of NLG texts on real users (e.g. testing reading time, or task completion) collaboration with psychologists will enable us to broaden and deepen this strand, looking at more fine-grained measures of reader behaviour (eg using eye-tracking) and assessments of a wider range of effects (such as emotional impact). In general, NLG can offer to psychologists the opportunity to further formalise and test their theories in more realistic settings. In return, results from psychology can inform our user and context models, as well as providing evidence about the effects of language alternatives in controlled settings.
  2. User Modelling and Affective computing. Affective computing is computing that relates to, arises from, or deliberately influences emotions or other non-strictly rational aspects of humans [13]. So far, however, work in "affective NLG" has aimed mainly to produce text that portrays the emotions of the writer, rather than considering how linguistic factors can affect the emotions of the reader. Work in affective computing may provide useful ways of formalising theories of emotion [10], modelling affective state and measuring effects on this state. In general, affective results may be easiest to monitor and achieve in multimodal communication systems, and this may require us to work with areas such as machine vision.
     
  3. NLG Architectures. The above issues (non-informative texts, reader variation), expose deficiencies in current NLG practices. Complex effects often involve a number of very different aspects of the text (e.g. sentence structuring, choice of vocabulary), interacting in non-trivial ways, and independent of the core factual content. Also, many effects arise from purely surface phenomena (eg text length, choice of words, word co-occurrences), and yet pipeline NLG architectures [17] discover surface effects only after all central decisions have been made. Abstract stylistic goals may have to be balanced against basic communicative tasks [21]; the COGENT project addresses some of these issues. There are a number of approaches to these problems: intelligent backtracking [4], 'overgeneration' architectures [5], and stochastic search [7], but such methods go beyond most current NLG architectures [8], and are still relatively untested on realistic examples.

 

Benefits

This research can be expected to have large benefits for both science and technology. From a scientific perspective, it will lead to theoretical results about some very poorly understood aspects of language. From an engineering point of view, it will establish practical methodologies for NLG development and evaluation. From a technological perspective, our work could lead to systems that help people in numerous ways, e.g. encouraging people to change their behaviour (cf. STOP, SkillSum), teaching children and other learners (cf. STANDUP), assisting specialists to understand complex data (cf. SumTime, BabyTalk). NLG research is on the cusp of a movement from simple informative software to more general, powerful and varied communication systems. Key to this development is a better understanding of how to affect people with natural language.

 

Bibliography

  1. Cawsey, A., Jones, R.B., and Pearson, J., "The Evaluation of a Personalised Information System for Patients with Cancer". User Modeling and User-Adapted Interaction, vol 10, no 1, 2000.
  2. Garrod, S. and Pickering, M., "Why is conversation so easy?". Trends Cogn Sciences8(1), pp 8-11, 2004.
  3. Hovy, E.H., "Pragmatics and Natural Language Generation". Artificial Intelligence43(2) pp153-198, 1990.
  4. Kamal, H. and Mellish, C., "An ATMS Approach to Systemic Sentence Generation". Procs of the Third International Conference on Natural Language Generation (INLG-04), New Forest, UK, pp 80-89, 2004.
  5. Langkilde, I. and Knight, K., "Generation that Exploits Corpus-based Statistical Knowledge". Procs of COLING/ACL, 1998.
  6. Law, A., Freer, Y., Hunter, J., Logie, R., McIntosh, N. and Quinn, J., "A Comparison of Graphical and Textual Presentations of Time Series Data to Support Medical Decision Making in the Neonatal Intensive Care Unit". Jnl of Clinical Monitoring and Computing, to appear (2005).
  7. Manurung, H., Ritchie, G., and Thompson, H., "A flexible integrated architecture for generating poetic texts". Procs of the Fourth Symposium on Natural Language Processing (SNLP 2000), Chiang Mai, Thailand, May 2000.
  8. Mellish, C. and Evans, R., "Implementation Architectures for Natural Language Generation". Natural Language Engineering, 10(3/4): pp 261-282, 2004.
  9. Moore, J., Porayska-Pomsta, K., Varges, S. and Zinn, C., "Generating Tutorial Feedback with Affect". Procs of the Seventeenth International Florida Artificial Intelligence Research Symposium Conference (FLAIRS), AAAI Press, 2004.
  10. Oatley, K. and Jenkins, J., Understanding Emotions, Blackwell, 1996.
  11. Oberlander, J. and Gill, A., "Individual differences and implicit language: Personality, parts-of-speech and pervasiveness." In Procs of the 26th Annual Conference of the Cognitive Science Society, pp1035-1040. Chicago, August 5-7, 2004.
  12. O'Donnell, M., Knott, A., Mellish, C. and Oberlander, J., "ILEX: The Architecture of a Dynamic Hypertext Generation System". Natural Language Engineering, 7: pp 225-250, 2001.
  13. Picard, R. W., Affective Computing. MIT Press, 1997.
  14. Piwek, P., "An Annotated Bibliography of Affective Natural Language Generation". Version 1.3 available from http://www.itri.brighton.ac.uk/~Paul.Piwek/topic-papers.html
  15. Rayner, K. and Pollatsek, A., The Psychology of Reading, Lawrence Erlbaum Associates, 1995.
  16. Reed, C. and Norman, T.J. (eds), Argumentation Machines: New Frontiers in Argumentation and Computation. Dordrecht: Kluwer, 2004.
  17. Reiter, E. and Dale, R., Building Natural Language Generation Systems. Cambridge: CUP, 2000.
  18. Reiter, E., Sripada, S., Hunter, J., Yu J., Davy I., "Choosing Words in Computer-Generated Weather Forecasts". Artificial Intelligence167(1-2): pp 137-169, 2005.
  19. Ritchie, G., "Current directions in computational humour". Artificial Intelligence Review16(2): pp 119-135, 2001.
  20. de Rosis, F. and Grasso, F., "Affective Natural Language Generation". In A. Paiva (ed.), Affective Interactions, Springer LNAI 1814, 2000.
  21. van Deemter, K., "Is Optimality-Theoretic Semantics Relevant for NLP?". Jnl of Semantics21(3), 2004.
  22. Walker, Marilyn A., Cahn, Janet E. and Whittaker, Stephen J., "Improvising linguistic style: social and affective bases for agent personality". Pp. 96 - 105 in Proc. 1st International Conference on Autonomous Agents, Marina del Rey, USA, 1997.
  23. Walker, M., Whittaker, S., Stent, A., Maloor, P., Moore, J., Johnston, M., Vasireddy, G. "Generation and Evaluation of User Tailored Responses in Multimodal Dialogue". Cognitive Science, 28(5), pp 811-840, 2003.
  24. Zukerman, I. and Litman, D. "Natural Language Processing and User Modeling: Synergies and Limitations". User Modeling and User-Adapted Interaction, 11(1-2), pp 129 - 158, 2001.
  25. Zinn, C., Moore, J. and Core, M., "Multimodal Intelligent Information Presentation". O. Stock and M. Zancanaro (eds.), Text, Speech and Language Technology, Vol. 27, pages 227-254, Kluwer Academic Publishers, 2005 (in press).