MixCSE:Unsupervised Sentence Representation via Contrastive Learning with MixingNegatives

asd8705

已于 2022-04-25 09:53:27 修改

阅读量652

点赞数

分类专栏：自然语言处理文章标签：自然语言处理

于 2022-04-20 19:50:38 首次发布

本文链接：https://blog.csdn.net/asd8705/article/details/124305514

版权

自然语言处理专栏收录该内容

19 篇文章

订阅专栏

本文介绍了一种基于GitHub开源项目MixCSE_AAAI2022的无监督句子表示学习方法，通过对比学习和混合负样本策略，提高了模型在多个语料库上的STS和MR等任务上的性能。研究显示，这种方法在STS12平均得分上达到76.48%，在多个NLP基准上展示了优势。

摘要生成于 C知道，由 DeepSeek-R1 满血版支持，前往体验 >

代码地址:GitHub - BDBC-KG-NLP/MixCSE_AAAI2022: Code for AAAI 2022 paper Unsupervised Sentence Representation via Contrastive Learning with Mixing Negatives

------ test ------
+-------+-------+-------+-------+-------+--------------+-----------------+-------+
| STS12 | STS13 | STS14 | STS15 | STS16 | STSBenchmark | SICKRelatedness | Avg. |
+-------+-------+-------+-------+-------+--------------+-----------------+-------+
| 69.87 | 82.36 | 74.48 | 82.20 | 78.21 | 77.84 | 70.43 | 76.48 |
+-------+-------+-------+-------+-------+--------------+-----------------+-------+
+-------+-------+-------+-------+-------+-------+-------+-------+
| MR | CR | SUBJ | MPQA | SST2 | TREC | MRPC | Avg. |
+-------+-------+-------+-------+-------+-------+-------+-------+
| 80.33 | 85.88 | 94.91 | 88.95 | 84.52 | 83.47 | 76.15 | 84.89 |
+-------+-------+-------+-------+-------+-------+-------+-------+