2021
Ce document est lié à :
info:eu-repo/semantics/altIdentifier/doi/10.18627/jslg.37.3.202111.335
http://creativecommons.org/licenses/by-nc/ , info:eu-repo/semantics/OpenAccess
Jeong-Woo Yoon et al., « A Methodology of Constructing Sentiment-Annotated Datasets for Training a FbSA model », HAL-SHS : linguistique, ID : 10.18627/jslg.37.3.202111.335
We report the construction of the FeSAD (Feature Sentiment Annotated) dataset based on the Semi-automatic Symbolic Propagation (SSP) method for FbSA (Feature-based Sentiment Analysis) in Korean. FeSAD was constructed in a 2-step annotation process: the SSP method was applied first, and then human annotators revised the annotated results. The linguistic resources for SSP consist of LGG (Local Grammar Graph) patterns and the DECO (Dictionnaire Electronique du COreen) Korean machine-readable dictionaries. We evaluated the performance of the SSP-based approach with texts in the cosmetics and food domains. The results reach 0.93 and 0.90 F1-score for these domains, which shows that the SSP-based approach could reduce the amount of the human annotators’ task to 10%.