Skip to Main content Skip to Navigation
Master Thesis

Étude du contrôle de contenu et de la compression de la génération automatique neuronale de textes

Abstract : Data-to-text generation is one of the most popular tasks in NLP. Data-to-text generation models can be divided into three categories : the first one is based on templates, the second is end-to-end models which can generate texts directly from data without any intermediate stage and the last one is pipeline models. Most data-to-text generation systems work on improving text fluency and grammatical correctness, disregarding control over text structure and length. However, control plays an important part in industrial NLG applications such as advertisements and product descriptions.
In this paper, we first introduce the task of data-to-text generation and the datasets often used in the task. Then we investigate state-of-the-art models, which takes up a big part of the paper. Finally, we apply different strategies of control in data-to-text generation systems particularly from the aspects of text structure and text length. Regarding text structure, we present an approach of alignment between input and target. It makes sure that the sentence count, the entity order and the distribution of input entites into sentences in both sides are the same. As for control over text length, we show two different approaches. One is to supply length constraint as input while the other is to force the end-of-sentence tag to be included at each step when using top-k decoding strategy. We propose four metrics to assess the degree to which these methods will affect a NLG system’s ability to control text structure and length. Our quantitative analyses demonstrate that all the methods enhance efficiently the system’s ability with a slight decrease in text fluency.
Document type :
Master Thesis
Complete list of metadatas

Cited literature [75 references]  Display  Hide  Download

https://dumas.ccsd.cnrs.fr/dumas-02972191
Contributor : Uga - Bulles <>
Submitted on : Tuesday, October 20, 2020 - 11:16:12 AM
Last modification on : Tuesday, November 24, 2020 - 4:00:19 PM

File

LENG_Yuanmin_M2 IDL.pdf
Files produced by the author(s)

Identifiers

  • HAL Id : dumas-02972191, version 1

Citation

Yuanmin Leng. Étude du contrôle de contenu et de la compression de la génération automatique neuronale de textes. Sciences de l'Homme et Société. 2020. ⟨dumas-02972191⟩

Share

Metrics

Record views

22

Files downloads

16