Advanced search
Start date
Betweenand
Related content
(Reference retrieved automatically from Web of Science through information on FAPESP grant and its corresponding number as mentioned in the publication by the authors.)

Multi-Domain Aspect Extraction Using Bidirectional Encoder Representations From Transformers

Full text
Author(s):
Santos, Brucce Neves Dos [1] ; Marcacini, Ricardo Marcondes [1] ; Rezende, Solange Oliveira [1]
Total Authors: 3
Affiliation:
[1] Univ Sao Paulo, Inst Math & Comp Sci, BR-13566590 Sao Carlos - Brazil
Total Affiliations: 1
Document type: Journal article
Source: IEEE ACCESS; v. 9, p. 91604-91613, 2021.
Web of Science Citations: 0
Abstract

Deep learning and neural language models have obtained state-of-the-art results in aspects extraction tasks, in which the objective is to automatically extract characteristics of products and services that are the target of consumer opinion. However, these methods require a large amount of labeled data to achieve such results. Since data labeling is a costly task, there are no labeled data available for all domains. In this paper, we propose an approach for aspect extraction in a multi-domain transfer learning scenario, thereby leveraging labeled data from different source domains to extract aspects of a new unlabeled target domain. Our approach, called MDAE-BERT (Multi-Domain Aspect Extraction using Bidirectional Encoder Representations from Transformers), explores neural language models to deal with two major challenges in multi-domain learning: (1) inconsistency of aspects from target and source domains and (2) context-based semantic distance between ambiguous aspects. We evaluated our MDAE-BERT considering two perspectives (1) the aspect extraction performance using F1-Macro and Accuracy measures; and (2) by comparing the multi-domain aspect extraction models and single-domain models for aspect extraction. In the first perspective, our method outperforms the LSTM-based approach. In the second perspective, our approach proved to be a competitive alternative compared to the single-domain model trained in a specific domain, even in the absence of labeled data from the target domain. (AU)

FAPESP's process: 19/25010-5 - Semantically enriched representations for Portuguese textmining: models and applications
Grantee:Solange Oliveira Rezende
Support Opportunities: Regular Research Grants
FAPESP's process: 19/07665-4 - Center for Artificial Intelligence
Grantee:Fabio Gagliardi Cozman
Support Opportunities: Research Grants - Research Program in eScience and Data Science - Research Centers in Engineering Program