Instance-based natural language generation

Abstract

We investigate the use of instance-based ranking methods for surface realization in natural language generation. Our approach to instance-based natural language generation (IBNLG) employs two components: a rule system that 'overgenerates' a number of realization candidates from a meaning representation and an instance-based ranker that scores the candidates according to their similarity to examples taken from a training corpus. We develop an efficient search technique for identifying the optimal candidate based on a novel extension of the A * algorithm. The rule system is produced automatically from a semantically annotated fragment of the Penn Treebank II containing management succession texts. We detail the annotation scheme and grammar induction algorithm and evaluate the efficiency and output of the generator. We also discuss issues such as input coverage (completeness) and fluency that are relevant to surface generation in general.

Stochastic Language Generation in Dialogue using Factored Language Models

“…Factored language models have been used for surface realization within the OpenCCG framework (White, Rajkumar, and Martin 2007;Espinosa, White, and Mehay 2008). More generally, chart generators for different grammatical formalisms have been trained from syntactic treebanks (Nakanishi, Miyao, & Tsujii 2005;Cahill and van Genabith 2006;White, Rajkumar, and Martin 2007), as well as from semantically annotated treebanks (Varges and Mellish 2001). Because manual syntactic annotation is costly and syntactic parsers do not necessarily perform well at labeling spoken language utterances, the present work focuses on the generation of surface forms directly from semantic concepts.…”

On the role of linguistic descriptions of data in the building of natural language generation systems

“…For instance, another interesting convergence point between NLG and LDD is related to hybrid rule-based/machine learning approaches in NLG, which combine rule-based overgeneration of candidate texts with ranking based on machine learning techniques (e.g., the proposal by Varges and Mellish [31], described in Section 2.1). This kind of approaches resemble standard LDD algorithms which generate every possible candidate description (due to imprecision handling) and provide the fittest one according to several evaluation criteria.…”

“…Another interesting approach is given by Varges and Mellish in [31], who propose an overgeneration-and-ranking approach which generates many possible candidate output sentences through a rule-based grammar and then selects the fittest one. Gkatzia et al present in [32] a methodology that treats content selection as a multi-label classification problem. This approach was applied to the generation of student feedback reports based on data for several factors.…”

Survey of the State of the Art in Natural Language Generation: Core tasks, applications and evaluation

“…Generation decisions are taken using the instance-based KStar algorithm, which is shown to outperform a majority baseline on all classification decisions. Instancebased approaches to nlg are also discussed by Varges and Mellish (2010), albeit in an overgenerate-and-rank approach where rules overgenerate candidates, which are then ranked by comparison to the instance base.…”