Mitigating Contradictions in Dialogue Based on Contrastive Learning
32582Authors
nan
Abstract
AbstractChatbot models have achieved remarkable progress in recent years but tend to yield contradictory responses. In this paper, we exploit the advantage of contrastive learning technique to mitigate this issue. To endow the model with the ability of discriminating contradictory patterns, we minimize the similarity between the target response and contradiction related negative example. The negative example is generated with learnable latent noise, which receives contradiction related feedback from the pretrained critic. Experimental results show that our method helps to avoid contradictions in response generation while preserving response fluency, outperforming existing methods on both automatic and human evaluation.
Keywords
Full Content
Findings of the Association for Computational Linguistics: ACL 2022, pages 239 - 245 May 22-27, 2022 c⃝2022 Association for Computational Linguistics RuCCoN: Clinical Concept Normalization in Russian Aleksandr Nesterov1, Galina Zubkova1, Zulfat Miftahutdinov2, Vladimir Kokh1, Elena Tutubalina2,3,4, Artem Shelmanov5,7, Anton M. Alekseev6, Manvel Avetisian1, Andrey Chertok4,5, Sergey Nikolenko6,7,8 1 Sber AI Lab, Moscow, Russia 2 Kazan Federal University, Kazan, Russia 3 HSE University, Moscow, Russia 4 Sber AI, Moscow, Russia 5 AIRI, Moscow, Russia 6 St. Petersburg Department of Steklov Mathematical Institute, St. Petersburg, Russia 7 ISP RAS Research Center for Trusted Artificial Intelligence, Moscow, Russia 8 Neuromation OU, Tallinn, Estonia {AAlNesterov,GVZubkova}@sberbank.ru Abstract We present RuCCoN, a new dataset for clin- ical concept normalization in Russian manu- ally annotated by medical professionals. contains over 16,028 entity mentions manu-It ally linked to over 2,409 unique concepts from the Russian language part of the UMLS on- tology. We provide train/test splits for dif- ferent settings (stratified, zero-shot, and CUI- less) and present strong baselines obtained with state-of-the-art models such as SapBERT. At present, Russian medical NLP is lacking in both datasets and trained models, and we view this work as an important step towards filling this gap. Our dataset and annotation guidelines are available at https://github.com/ sberbank-ai-lab/RuCCoN. 1 Introduction Electronic health records and other clinical texts contain patient histories through the progression of diseases and represent a treasure trove of in- formation for medical specialists. This infor- mation is often unstructured, concealed in free- form text, which leads to the need for natural lan- guage processing on medical texts. Mentions of diseases, symptoms, drugs, and other concepts are highly variable, and since the medical vo- cabulary is very large, entity linking and con- cept normalization become hard and important problems. State-of-the-art models are increas- ingly successful in high-resource languages such as English or Spanish, where labeled datasets in- clude ShARe/CLEF eHealth 2013 Task 1 (Suomi- nen et al., 2013), SemEval-2014 Task 7 (Prad- han et al., 2014), SemEval-2015 Task 14 (El- hadad et al., 2015), MCN (Luo et al., 2019), CANTEMIST (Miranda-Escalada et al., 2020a), CodiEsp (Miranda-Escalada et al., 2020b), and oth- ers. However, little has been done for medical entity linking in many languages that are high- resource in other regards. One example is Rus- sian: it is among top 10 languages in the world and has many NLP datasets and resources, but the medical part of Russian NLP is underdevel- oped. The Russian UMLS includes translations of Medical Dictionary for Regulatory Activities (MedDRA) (Brown et al., 1999), Logical Observa- tion Identifiers Names and Codes (LOINC) (For- rey et al., 1996), and Medical Subject Headings (MeSH) (Coletti and Bleich, 2001), but it still only amounts to 1.8% of the English UMLS in vocabu- lary and 1.36% in source counts (NIH 2021, a). In this work, we present RuCCoN (Russian Clinical Concept Normalization), a new labeled dataset for clinical concept normalization in Rus- sian. We have employed medical professionals to label the dataset based on concepts from the Rus- sian UMLS (Section …
Content truncated to 500 words
Labeled Text
AbstractChatbot models have achieved remarkable progress in recent years but tend to yield contradictory responses. In this paper, we exploit the advantage of contrastive learning technique to mitigate this issue. To endow the model with the ability of discriminating contradictory patterns, we minimize the similarity between the target response and contradiction related negative example. The negative example is generated with learnable latent noise, which receives contradiction related feedback from the pretrained critic. Experimental results show that our method helps to avoid contradictions in response generation while preserving response fluency, outperforming existing methods on both automatic and human evaluation.
Publication Details
- Paper ID:
- 32582
- Year:
- 2022
- Publication:
- TIPSTER
- Publication Type:
- conf
- DOI:
- nan