Financial news is an indispensable source for both investors and regulators to conduct research and investment decisions. To focus on specific areas of interest among the massive financial news, there is an urgent necessity of automatic financial news annotation, which faces two challenges: (1) supervised data scarcity for sub-divided financial fields; (2) the multifaceted nature of financial news. To address these challenges, we target the automatic financial news annotation problem as a weakly-supervised hierarchical multi-label classification. We propose a method that needs no manual labeled data, but a label hierarchy with one keyword for each leaf label as supervision. Our method consists of three components: word embedding with heterogeneous information, multi-label pseudo documents generation, and hierarchical multi-label classifier training. Experimental results on data from a well-known Chinese financial news website demonstrate the superiority of our proposed method over existing methods.
展开▼
