Trainable Methods for Surface Natural Language Generation

TL;DR

This paper introduces three trainable surface natural language generation systems that produce grammatical phrases from semantic representations, utilizing annotated corpora with varying levels of syntactic information.

Contribution

It presents novel trainable NLG systems that leverage maximum entropy models and semantic annotations, advancing the ability to generate natural language phrases from domain-specific data.

Findings

NLG2 and NLG3 learn word choice and order effectively.

All systems successfully generate grammatical phrases.

Experiments demonstrate applicability in the air travel domain.

Abstract

We present three systems for surface natural language generation that are trainable from annotated corpora. The first two systems, called NLG1 and NLG2, require a corpus marked only with domain-specific semantic attributes, while the last system, called NLG3, requires a corpus marked with both semantic attributes and syntactic dependency information. All systems attempt to produce a grammatical natural language phrase from a domain-specific semantic representation. NLG1 serves a baseline system and uses phrase frequencies to generate a whole phrase in one step, while NLG2 and NLG3 use maximum entropy probability models to individually generate each word in the phrase. The systems NLG2 and NLG3 learn to determine both the word choice and the word order of the phrase. We present experiments in which we generate phrases to describe flights in the air travel domain.