Synonyms
Definition
Text generation is a subfield of natural language processing. It leverages knowledge in computational linguistics and artificial intelligence to automatically generate natural language texts, which can satisfy certain communicative requirements.
Historical Background
Research work in the text generation field first appeared in the 1970s. Goldman's work on natural language generation from a deep conceptual base appeared in [2]. In the 1980s, more significant work was contributed in this field: McDonald saw text generation as a decision making problem [6], Appelt on language planning (1981), McKeown [8]. In the 1990s, a generic architecture for text generation was discussed, Reiter [10], Hovy [3]. Still today, variations on the generic architecture is a still a widely discussed question, Mellish et al. [9].
Foundations
Text Generation, or Natural language generation (NLG), is usually compared with another subfield of natural language processing – natural language understanding (NLU), which is generally considered as the inverse process of the former. Because in a highly abstract level, NLG task synthesizes machine representation of information into natural language texts, while NLU task parses and maps natural language texts into machine representations. However, upon inspection at a more concrete level, they can hardly be seen as “opposite,” because they are very different in problem sets, and by internal representations.
Text Generation System Architecture
Input and Output
-
The input of text generation system is information represented in non-linguistic format, such as numerical, symbolical, graphical, etc. The output is understandable natural language in text format, such as messages, documents, reports, etc.
Architectures
The Generic Architecture
Despite difference in application backgrounds and realization details, many of the current text generation systems followed a general architecture, which is known as the Pipelined Architecture or Consensus Architecture, usually described as in Fig. 1([11]; Edward Hovy also had a similar representation for this architecture).
As seen in the Fig. 1, the “Pipelined Architecture” describes a general strategy of tackling text generation problem from macro to micro, from inner structure organization to outer surface realization. Thus, language components such as paragraphs, sentences, and words will be coherently arranged together to meet certain communicative requirements.
The following are the detailed descriptions of the above stages:
Stage 1: Document Planning
Also known as Text Planning, Discourse Planning or Macro Planning). This includes:
-
Content determination: Also know as content selection and organization, which is to discover and determine the major topics the text should cover, given a set of communicative goals and representations of information or knowledge.
-
Document structuring: Determining the overall structure of the text/document. This structure categorizes and organizes sentence-leveled language components into clusters. The relationship between different components inside a cluster can be explanatory, descriptive, comparative, causal, sequential, etc.
Stage 2: Micro Planning
Also know as Sentence Planning. This is to convert a document plan into a sequence of sentence or phrase specifications, including:
-
Aggregation: To combine several linguistic structures (e.g., sentences, paragraphs) into a single and coherent structure. An example: Tomorrow will be cold. Tomorrow will be windy.
-
→Tomorrow will be cold and windy.
-
-
Lexicalization: To choose appropriate words from possible lexicalizations based on the communicative background. Examples: (i) buy, purchase, take, etc, (ii) a lot of, large amounts of, etc.
-
Referring expression generation: To choose or introduce different means of reference for sentences, such as pronouns (pronominalization). There is usually more than one way to identify a specific object, for example: “Shakespeare, ” “the poet and playwright, ” “the Englishman, ”and “he/him” can all point to the same object. Example: Andrew wanted to sing at the birthday party.
-
→He wanted to sing at the birthday party.
-
→The boy wanted to sing at the birthday party.
-
Stage 3: Surface realization
Also know as Speech Synthesis. This is to finally synthesize the text according the text specifications made in the previous stages.
-
Structure realization: To mark up the text's surface structure, such as an empty line, or the boundaries between paragraphs, etc.
-
Linguistic realization: To smooth the text by inserting function words, reorder word sequences, and select appropriate inflections and tenses of words, etc.
Other Architectures:
Although the Pipelined Architecture provides a considerably articulate routine for text generation, it also provides predetermined restrictions for each stage in the process. Thus, the flexibility it can provide is limited, and is especially true for those sub-tasks in micro planning and surface realization stages. For example, the need for lexical selection can happen at any stage of the process. Thus, variations of the generic architecture and other methodologies have been discussed by many researchers (a recent discussion, Chris Mellish et al. [9]).
Key Applications
-
1.
Routine documentation or information generation: examples of information are weather forecast descriptions, transportation schedules, accounting spreadsheets, expert system knowledge bases, etc. Examples of documentation are technical reports and manuals, business letters, medical records, doctor prescriptions, etc.
-
2.
Literary writing: such as stories, poems, lyrics, couplets, etc. (Chinese couplet writer: generating a couplet sentence according to a given one. http://duilian.msra.cn).
Recommended Reading
Dale R. Introduction to the special issue on natural language generation. Comput. Linguistics, 24 (3):346–353, 1998.
Goldman N.M. Computer Generation of Natural Language from a Deep Conceptual Base. Ph.D. thesis, Stanford University, CA, 1974.
Hovy E.H. 1997, G.B.Varile, A. Zampolli (eds.). Language generation, Chapter 4. In Survey of the State of the Art in Human Language Technology, Cambridge University Press, Cambridge, pp. 139–163.
Hovy E.H. Natural language generation. Entry for MIT Encyclopedia of Computer Science. MIT Press, Cambridge, MA, pp.585–5881998,
Hovy E.H. Language generation. Entry for Encyclopedia of Cognitive Science, article 86. McMillan, London, 2000.
McDonald D.D. Natural Language Production as a Process of Decision Making Under Constraint. Ph.D. thesis, MIT Artificial Intelligence Laboratory, Cambridge, MA, 1980.
McDonald D.D. 2000, Dale, R. H. Moisl, H. (eds.). Somers 1Natural language generation, Chapter 7. In Handbook of Natural Language Processing, Marcel Dekker, New York, NY, pp. 147–180.
McKeown K.R. Text Generation: Using Discourse Strategies and Focus Constraints to Generate Natural Language Text. Cambridge University Press, Cambridge, 1985.
Mellish C, et al. A reference architecture for natural language generation systems. Nat. Lang. Eng., 12(1):1–34, 2006.
Reiter E. Has a consensus NL generation architecture appeared and is it psycholinguistically plausible? In Proc. 7th Int. Conf. on Natural Language Generation, pp. 163–170.1994,
Reiter E. and Dale R. Building Natural Language Generation Systems. Cambridge University Press, Cambridge, 2000.
Author information
Authors and Affiliations
Editor information
Editors and Affiliations
Rights and permissions
Copyright information
© 2009 Springer Science+Business Media, LLC
About this entry
Cite this entry
Zhang, L., Sun, JT. (2009). Text Generation. In: LIU, L., ÖZSU, M.T. (eds) Encyclopedia of Database Systems. Springer, Boston, MA. https://doi.org/10.1007/978-0-387-39940-9_416
Download citation
DOI: https://doi.org/10.1007/978-0-387-39940-9_416
Publisher Name: Springer, Boston, MA
Print ISBN: 978-0-387-35544-3
Online ISBN: 978-0-387-39940-9
eBook Packages: Computer ScienceReference Module Computer Science and Engineering