关系抽取 (Relation Extraction) 论文中常用的数据集:SemEval-2010 Task-8
Multi-Way Classification of Semantic Relations Between Pairs of Nominals
官方网址: http://semeval2.fbk.eu/semeval2.php
文档: https://www.aclweb.org/anthology/S10-1006/
数据: https://github.com/CrazilyCode/SemEval2010-Task8
BibTeX:
@inproceedings{hendrickx-etal-2010-semeval,
title = "{S}em{E}val-2010 Task 8: Multi-Way Classification of Semantic Relations between Pairs of Nominals",
author = "Hendrickx, Iris and
Kim, Su Nam and
Kozareva, Zornitsa and
Nakov, Preslav and
{\'O} S{\'e}aghdha, Diarmuid and
Pad{\'o}, Sebastian and
Pennacchiotti, Marco and
Romano, Lorenza and
Szpakowicz, Stan",
booktitle = "Proceedings of the 5th International Workshop on Semantic Evaluation",
month = jul,
year = "2010",
address = "Uppsala, Sweden",
publisher = "Association for Computational Linguistics",
url = "https://www.aclweb.org/anthology/S10-1006",
pages = "33--38",
}
示例数据:
句子:The <e1>fire</e1> inside WTC was caused by exploding <e2>fuel</e2>.
关系:Cause-Effect(e2,e1)
训练集:共含8000条数据
type | number | rate |
---|---|---|
Other | 1410 | 17.63% |
Cause-Effect | 1003 | 12.54% |
Component-Whole | 941 | 11.76% |
Entity-Destination | 845 | 10.56% |
Product-Producer | 717 | 8.96% |
Entity-Origin | 716 | 8.95% |
Member-Collection | 690 | 8.63% |
Message-Topic | 634 | 7.92% |
Content-Container | 540 | 6.75% |
Instrument-Agency | 504 | 6.30% |
测试集:共含2717条数据
type | number | rate |
---|---|---|
Other | 454 | 16.71% |
Cause-Effect | 328 | 12.07% |
Component-Whole | 312 | 11.48% |
Entity-Destination | 292 | 10.75% |
Message-Topic | 261 | 9.61% |
Entity-Origin | 258 | 9.50% |
Member-Collection | 233 | 8.58% |
Product-Producer | 231 | 8.50% |
Content-Container | 192 | 7.07% |
Instrument-Agency | 156 | 5.74% |
测评:官方给出了perl语言的测评程序
The scorer calculates and outputs the following statistics:
(1) confusion matrix, which shows
- the sums for each row/column: -SUM-
- the number of skipped examples: skip
- the number of examples with correct relation, but wrong directionality: xDIRx
- the number of examples in the answer key file: ACTUAL ( = -SUM- + skip + xDIRx )
(2) accuracy and coverage
(3) precision P, recall R, and F1-score for each relation
(4) micro-averaged P, R, F1, where the calculations ignore the Other category.
(5) macro-averaged P, R, F1, where the calculations ignore the Other category.
Note that in scores (4) and (5), skipped examples are equivalent to those classified as Other.
So are examples classified as relations that do not exist in the key file (which is probably not optimal).
The scoring is done three times:
(i) as a (2*9+1)-way classification
(ii) as a (9+1)-way classification, with directionality ignored
(iii) as a (9+1)-way classification, with directionality taken into account.
The official score is the macro-averaged F1-score for (iii).
相关论文:
论文 | 结果 | 备注 |
---|---|---|
(2014COLING) Relation Classification via Convolutional Deep Neural Network | 82.7% | 预训练好的词向量表示+WordNet |
(2015ACL) Classifying Relations by Ranking with Convolutional Neural Networks | 84.1% | 预训练好的词向量表示 |
持续更新中。。。