TATA: Stance Detection via Topic-Agnostic and Topic-Aware Embeddings

意见态度检测对于理解互联网上不同的态度和信仰非常重要。然而，考虑到一段话对于一个特定主题的态度往往高度依赖于该主题，建立一个可以推广到未知主题的态度检测模型是困难的。在这项工作中，我们提出使用对比学习以及一个包含各种不同主题的未标记新闻文章数据集来训练面向主题无关（TAG）和面向主题感知（TAW）的嵌入，以用于下游的态度检测。将这些嵌入结合在我们的完整TATA模型中，我们在几个公共态度检测数据集上实现了最先进的性能（在Zero-shot VAST数据集上的0.771 F1分数）。我们在https://github.com/hanshanley/tata发布了我们的代码和数据。

The paper aims to solve the problem of building a stance detection model that generalizes to unseen topics by proposing the use of contrastive learning and an unlabeled dataset of news articles to train topic-agnostic and topic-aware embeddings.

The key idea of the paper is to use contrastive learning and an unlabeled dataset to train embeddings that are both topic-agnostic and topic-aware, and then combine these embeddings in the TATA model for downstream stance detection.

The paper achieves state-of-the-art performance on several public stance detection datasets, with a 0.771 F1-score on the Zero-shot VAST dataset. The code and data are released on GitHub. The paper also discusses the limitations of the proposed approach and suggests potential future directions for research.

Related work in this field includes 'A Hierarchical Multi-task Approach for Learning Embeddings from Semantic Tasks' by Zhang et al. and 'Learning to Attend On Essential Terms: An Enhanced Retriever-Reader Model for Domain-Specific Question Answering' by Wang et al.