An encoder–decoder approach to mine conditions for engineering textual data
Data engineering seeks to support artificial intelligence processes that extract knowledge from raw data. Many such data are rendered in natural language from which entity-relation extractors extract facts and opinion miners extract opinions; the goal of condition mining is to mine the conditions th...
| Autores: | , |
|---|---|
| Formato: | artículo |
| Estado: | Versión publicada |
| Fecha de publicación: | 2020 |
| País: | España |
| Recursos: | Universidad de Sevilla (US) |
| Repositorio: | idUS. Depósito de Investigación de la Universidad de Sevilla |
| OAI Identifier: | oai:idus.us.es:11441/143460 |
| Acesso em linha: | https://hdl.handle.net/11441/143460 https://doi.org/10.1016/j.engappai.2020.103568 |
| Access Level: | acceso abierto |
| Palavra-chave: | Condition mining Natural language processing Neural networks |
| Resumo: | Data engineering seeks to support artificial intelligence processes that extract knowledge from raw data. Many such data are rendered in natural language from which entity-relation extractors extract facts and opinion miners extract opinions; the goal of condition mining is to mine the conditions that have an influence on them. In this article, a new condition mining method is proposed. It relies on a deep neural network and attempts to overcome the limitations of existing methods for condition mining that we reviewed. The materials used include readily-available software components for natural language processing and a large multi-lingual, multi-topic dataset. The common information retrieval performance measures were used to assess the results, namely: precision, which is the fraction of correct conditions to the mined ones, recall, which is the fraction of correct conditions that have been mined to the total number of correct conditions, and the F1 score, which is the harmonic mean of precision and recall. The results of the experimental analysis prove that the new proposal can attain an F1 score that is significantly greater than with existing methods. Furthermore, a comprehensive analysis of the dataset was performed, which revealed two key findings: the connectives follows a long-tail distribution and the conditions are quite dissimilar from a semantic point of view. |
|---|