Shallow generation corresponds to shallow analysis in an interesting way. Whereas the latter ignores information by yielding incomplete analysis, the former adds information with help of output text not based on the input representation. For instance, in the domain of air quality measurements, the input " location: 'St. Avold' " may give rise to the output "at the measuring station of St. Avold", where "measuring station" is domain-specific.
Time-efficient shallow approaches to small applications can be very quickly developed by defining domain-specific grammars. Good examples are report generators (stock market, weather, air quality, limited dialogue, or NL access to database contents. They are less recommendable when the application requires a large variety of outputs and consequently a large linguistic knowledge base.
Note that the distinction between shallow and deep generation is gradual rather than bipolar.
- Smoking Termination with cOmputerised Personalisation (STOP)
- Transnational Environmental Management Support & Information System (TEMSIS)
- Natural Language Generation Group at the Department of Computing Science, University of Aberdeen (NLG Group Aberdeen)
- Language Technology Lab (LT Lab)
- University of Aberdeen
- Ehud Reiter
- Stephan Busemann
Building Natural Language Generation Systems.
Ehud Reiter and Robert Dale.
Cambridge University Press. Cambridge UK. 2000.
Real vs. Template-Based natural Language generation: A False Opposition?
Kees van Deemter, Emiel Krahmer and Mariet Theune.
Computational Linguistics 31 (1), 2005. 15-23.
A Flexible Shallow Approach to Text Generation.
Stephan Busemann and Helmut Horacek.
9th International Natural Language Generation Workshop (INLG '98), August 5-7.
Niagara-on-the-Lake, Ontario, Canada. 1998. 238-247.
Generierung natürlichsprachlicher Texte.
Günther Görz, Claus-Rainer Rollinger, Josef Schneeberger. Künstliche Intelligenz. Oldenburg Verlag, München, Wien. 2000. 783-814.