Automatic generation of Chinese classical poetry is still a challenging problem in artificial intelligence. Recently, Encoder-Decoder models have provided a few viable methods for poetry generation. However, by reviewing the prior methods, two major issues still need to be settled: 1) most of them are one-stage generation methods without further polishing; 2) they rarely take into consideration the restrictions of poetry, such as tone and rhyme. Intuitively, some ancient Chinese poets tended first to write a coarse poem underlying aesthetics and then deliberated its semantics; while others first create a semantic poem and then refine its aesthetics. On this basis, in order to better imitate the human creation procedure of poems, we propose a two-stage method (i.e., restricted polishing generation method) of which each stage focuses on the different aspects of poems (i.e., semantics and aesthetics), which can produce a higher quality of generated poems. In this way, the two-stage method develops into two symmetrical generation methods, the aesthetics-to-semantics method and the semantics-to-aesthetics method. In particular, we design a sampling method and a gate to formulate the tone and rhyme restrictions, which can further improve the rhythm of the generated poems. Experimental results demonstrate the superiority of our proposed two-stage method in both automatic evaluation metrics and human evaluation metrics compared with baselines, especially in yielding consistent improvements in tone and rhyme.

Intelligent Financial Advisors (IFAs) in online financial applications (apps) have brought new life to personal investment by providing appropriate and high-quality portfolios for users. In real-world scenarios, identifying potential clients is a crucial issue for IFAs, i.e., identifying users who are willing to purchase the portfolios. Thus, extracting useful information from various characteristics of users and further predicting their purchase inclination are urgent. However, two critical problems encountered in real practice make this prediction task challenging, i.e., sample selection bias and data sparsity. In this study, we formalize a potential conversion relationship, i.e., user $\mathrm{\to }$ activated user $\mathrm{\to }$ client and decompose this relationship into three related tasks. Then, we propose a Multitask Feature Extraction Model (MFEM), which can leverage useful information contained in these related tasks and learn them jointly, thereby solving the two problems simultaneously. In addition, we design a two-stage feature selection algorithm to select highly relevant user features efficiently and accurately from an incredibly huge number of user feature fields. Finally, we conduct extensive experiments on a real-world dataset provided by a famous fintech bank. Experimental results clearly demonstrate the effectiveness of MFEM.