SIGGEN: What is Text Generation?

The objective of natural language generation (NLG) or text generation systems is to produce coherent natural language texts which satisfy a set of one or more communicative goals. To achieve these goals, the generated text should be (among other things):

• coherent: using well-connected, sensible and comprehensible English;
• accurate: containing accurate information (or it could lead to the user making false inferences);
• valid: causing the user to make the desired inferences (for example, telling a naive user that the koala looks like a teddy bear and not telling her that it doesn't behave like one may result in a nasty surprise);
• informative: presenting new and interesting information to the user (for example, if a user who is unfamiliar with any marsupials requests a description of the kangaroo, then it is important that the description highlight the features of marsupials, and more specifically, the uniqueness of the kangaroo within this context; we should not produce a description which discusses some more abstract properties of the kangaroo, since the user first needs to form a mental model of what the entity is, and how it relates to other entities she is already familiar with);
• understandable: including information which the user can understand (for example, when describing the accordion to a person who is unfamiliar with it, we should: (i) use non-technical terms, since highly technical terms will lead to confusion; and (ii) take advantage of the hearer's existing knowledge of other entities such as the piano in order to make the description more understandable); and
• relevant: including information which is relevant to the current discourse goal and not redundant.

Figure 1 shows the traditional architecture of NLG systems. There are traditionally two main processing components: the text planning component and the surface realisation component. More recently, there has been a move towards including a third component, the sentence planning stage, between these two components. The text planning stage typically encapsulates all those decisions involving choices of what to say in a text; these texts are constructed based on an underlying representation of knowledge. Based on the discourse goals, the text planner must decide what is relevant in a particular situation, and organise this content in a way that allows realisation of a coherent discourse; this is called the CONTENT SELECTION and ORGANISATION stage. The text planning component achieves this by composing a discourse plan; there are several techniques for achieving this, for example, in McKeown's (1985) schema-based approach, several schemata (or discourse plan outlines) are defined, and may be instantiated from information contained in the knowledge base. On the other hand, Rhetorical Structure Theory (RST) identifies the coherence relations that exist between individual segments (of varying sizes) of a text and builds up a hierarchical structure of these relations. The reader is referred to Mann and Thompson (1987) and Hovy (1993) for more information.

Figure 1: Traditional NLG System architecture

A model of the user's knowledge can be used by an NLG system to tailor a text to the individual user's knowledge and needs. In addition, the ongoing discourse with a particular user can be recorded in the discourse history component to enable the system to adapt texts to what has been said before. Some approaches to adapting texts to a model of the hearer are described by Paris (1987) and Moore (1989).

The discourse plan is realised as natural language utterances by the surface realisation component. This makes use of knowledge of the natural language's grammar and lexicon to produce well-formed utterances that convey the required semantic content. The reader is referred to Reiter and Dale (2000) for a survey of approaches.

More Information

The following provide some introductory information about text generation for newcomers to the field:

• Dale & Reiter's ANLP-97 Tutorial on Building Applied Natural Language Generation Systems. (PPT7) (PPT4) (Postscript)
• The Survey of the State of the Art of Human Language Technology, was sponsored by the National Science Foundation and European Union.  It is a useful overview of the state of the art in many areas of language technology, including language generation (chapter 4). 
• CoGenTex's introduction to text generation.

References

Eduard H. Hovy. (1993) Automated discourse generation using discourse structure relations. Artificial Intelligence 63:341--385. Elsevier Science Publishers, Amsterdam.

Kathleen R. McKeown. (1985) Text Generation: Using Discourse Strategies and Focus Constraints to Generate Natural Language Text. Cambridge University Press.

William C. Mann and S. A. Thompson. (1987) Rhetorical structure theory: description and construction of text structures. In G. Kempen (ed.) Natural Language Generation. Martinus Nijhoff Publishers, pp. 85--95.

Johanna D. Moore. (1989) A Reactive Approach to Explanation in Expert and Advice-Giving Systems. Ph.D. Dissertation, University of California, Los Angeles.

Cecile L. Paris. (1987) The Use of Explicit User Models in Text Generation: Tailoring to a User's Level of Expertise. Ph.D. Dissertation, Columbia University.

Ehud Reiter and Robert Dale. (2000) Building Natural Language Generation Systems. Cambridge University Press.

SIGGEN