Neural natural language generation with unstructured contextual information

Abstract

[EU] Lan honetan, hizkuntza naturalaren sorrera automatikoan informazio ez-egituratuaren esplotazioak izan dezakeen eragina aztertzen da. Bere helburu nagusia, sistema batek aurrez ikusi gabeko informazioa erabiliz testu koherentea sortzeko duen gaitasuna ebaluatzea da. Corpus berri bat ere aurkezten da, zeregin honetarako bereziki prestatutako Amazon Review corpusaren aldaera bat, produktuen deskribapenak input gisa erabiliz, erabiltzaileen iritziak automatikoki sortzeko erabiltzen dena. Hainbat deep learning ereduk eginkizun honetan lortzen dituzten emaitzak konparatzen dira eta informazio ez egituratua ustiatzeko gaitasun maila ezberdina dutela erakusten da.

[EN] In this work, we present a novel task for automatic natural language generation, based on the exploitation of unstructured contextual information. The main aim of the task is to enable the evaluation of a system's capability to generate coherent text based on previously unseen and unstructured information. A new corpus was prepared specifically for the task, based on the Amazon Review corpus with product descriptions used as input for the generation of user reviews. Different deep learning generation models were implemented and compared under the proposed task, with significant differences in terms of their ability to exploit unstructured contextual information.