This paper introduces Astock, a comprehensive platform and dataset designed to systematically study natural language processing (NLP)-aided automated stock trading algorithms. Unlike previous work, Astock provides stock-specific financial news and various stock factors, enabling more realistic evaluation through financial-relevant metrics such as annualized rate of return and maximum drawdown. The platform facilitates the development and assessment of trading strategies by integrating textual and numerical data, addressing gaps in existing datasets that often lack specificity or comprehensive financial context.

Technically, the authors propose a novel method centered on semantic role labeling pooling (SRLP), which leverages semantic role labeling to create compact representations of news paragraphs. This approach is combined with other stock factors to enhance prediction accuracy for trading decisions. Additionally, a self-supervised learning strategy based on SRLP is introduced to improve the system's generalization to out-of-distribution data. Experimental results demonstrate that the proposed method outperforms all baselines in terms of annualized rate of return and achieves better maximum drawdown compared to the CSI300 and XIN9 indices in real trading scenarios. The Astock dataset and code are publicly available, contributing to the field by providing a robust resource for future research in NLP-driven financial decision-making.