FERGUS: Flexible Empiricist/Rationalist

Generation Using Syntax

What is FERGUS ?

FERGUS is a trainable, statistical natural language generation (NLG) component that can be used in spoken dialog and speech translation systems.

Natural language generation has been viewed to consist of the following components.

Text Planning: content and structure of the target text are determined to achieve the overall communicative goal.
Sentence Planning: linguistic means -- in particular, lexical and syntactic means -- are determined to convey smaller pieces of meaning.
Realization: the specification chosen in sentence planning is transformed into a surface string, by linearizing and inflecting words in the sentence .

At the present time, FERGUS consists of statistical models for surface realization and some aspects of sentence planning such as lexical choice and morphology.

Why FERGUS ?

NLG systems are being applied in restricted domains using carefully hand-crafted ("rationalist") approaches. These approaches generate language language that is highly controlled with guaranteed quality. However, the drawback of these approaches is that it is not only tedious and time consuming to develop but are also limited in terms of domain and language portability.

FERGUS is intended to overcome these limitations by combining the merits of rationalist approaches with the benefits of statistical ("empiricist") approaches. Statistical models allow FERGUS to be a trainable, portable and robust NLG system.

PUBLICATIONS

``John Chen, Srinivas Bangalore, Owen Rambow and Marilyn Walker, Towards Automatic Generation of Natural Language Generation Systems'', International Conference on Computational Linguistics (COLING 2002), Taipei, Taiwan, 2002.

``Srinivas Bangalore and Owen Rambow, Corpus-Based Lexical Choice in Natural Language Generation'', Association of Computational Linguistics (ACL 2000) , Hongkong, China, October 2000.

``Srinivas Bangalore and Owen Rambow, Exploiting a Probabilistic Hierarchical Model for Generation'', International Conference on Computational Linguistics (COLING 2000) , Saarbrucken, Germany, August 2000.

``Srinivas Bangalore, Owen Rambow, Steven Whittaker, Evaluation Metrics for Generation'', International Conference on Natural Language Generation (INLG 2000) , Mitzpe Ramon, Israel, July 2000.

``Srinivas Bangalore and Owen Rambow, Using TAG, a Tree Model, and a Language Model for Generation'', Fifth International Workshop on Tree-Adjoining Grammars (TAG+) , TALANA, Paris, May 2000.

PRESENTATIONS

Slides from presentation at Association of Computational Linguistics (ACL 2000) , Hongkong, China, October 2000.

Slides from presentation at International Conference on Computational Linguistics (COLING 2000) , Saarbrucken, Germany, August 2000.

Slides from presentation at International Conference on Natural Language Generation (INLG 2000) , Mitzpe Ramon, Israel, July 2000.