KCC 2021
Current Result Document :
ÇѱÛÁ¦¸ñ(Korean Title) |
¹®¼ ¼öÁØ °ü°è ¿¹ÃøÀ» À§ÇÑ ¼³¸í °¡´ÉÇÑ Áõ°Å ÃßÃ⠸𵨠|
¿µ¹®Á¦¸ñ(English Title) |
Evidence Retrieval toward an explainable model for Document Level Relation Extraction |
ÀúÀÚ(Author) |
¹èÇö°æ ÀÌȯÈñ À̹οì Á¤±³¹Î
Hyunkyung Bae
Hwanhee Lee
Minwoo Lee
Kyomin Jung
|
¿ø¹®¼ö·Ïó(Citation) |
VOL 48 NO. 01 PP. 0606 ~ 0608 (2021. 06) |
Çѱ۳»¿ë (Korean Abstract) |
|
¿µ¹®³»¿ë (English Abstract) |
Document-level relation extraction aims to extract relations among entities in a document. In this paper, we present the Evidence Retrieval for Relation Extraction (ER4RE) model that promotes the interpretability of the model prediction. ER4RE consists of an evidence retriever and a relation predictor, which work sequentially. The evidence retriever identifies the supporting sentences and computes the context representation for a given query. Then, the relation predictor takes the context information to infer the relations. We demonstrate that ER4RE achieves a significant performance improvement with oracle evidence sentences, which implies that if the model identifies supporting evidence for a given query, it can easily predict the relations. |
Å°¿öµå(Keyword) |
|
ÆÄÀÏ÷ºÎ |
PDF ´Ù¿î·Îµå
|