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.

Cognitive diagnosis is an important issue of intelligent education systems, which aims to estimate students’ proficiency on specific knowledge concepts. Most existing studies rely on the assumption of static student states and ignore the dynamics of proficiency in the learning process, which makes them unsuitable for online learning scenarios. In this paper, we propose a unified temporal item response theory (UTIRT) framework, incorporating temporality and randomness of proficiency evolving to get both accurate and interpretable diagnosis results. Specifically, we hypothesize that students’ proficiency varies as a Wiener process and describe a probabilistic graphical model in UTIRT to consider temporality and randomness factors. Furthermore, based on the relationship between student states and exercising answers, we hypothesize that the answering result at time k contributes most to inferring a student's proficiency at time k, which also reflects the temporality aspect and enables us to get analytical maximization (M-step) in the expectation maximization (EM) algorithm when estimating model parameters. Our UTIRT is a framework containing unified training and inferencing methods, and is general to cover several typical traditional models such as Item Response Theory (IRT), multidimensional IRT (MIRT), and temporal IRT (TIRT). Extensive experimental results on real-world datasets show the effectiveness of UTIRT and prove its superiority in leveraging temporality theoretically and practically over TIRT.

Generating molecules with desired properties is an important task in chemistry and pharmacy. An efficient method may have a positive impact on finding drugs to treat diseases like COVID-19. Data mining and artificial intelligence may be good ways to find an efficient method. Recently, both the generative models based on deep learning and the work based on genetic algorithms have made some progress in generating molecules and optimizing the molecule's properties. However, existing methods need to be improved in efficiency and performance. To solve these problems, we propose a method named the Chemical Genetic Algorithm for Large Molecular Space (CALM). Specifically, CALM employs a scalable and efficient molecular representation called molecular matrix. Then, we design corresponding crossover, mutation, and mask operators inspired by domain knowledge and previous studies. We apply our genetic algorithm to several tasks related to molecular property optimization and constraint molecular optimization. The results of these tasks show that our approach outperforms the other state-of-the-art deep learning and genetic algorithm methods, where the z tests performed on the results of several experiments show that our method is more than 99% likely to be significant. At the same time, based on the experimental results, we point out the insufficiency in the experimental evaluation standard which affects the fair evaluation of previous work.

Recent years have witnessed the prevalence of recommender systems in various fields, which provide a personalized recommendation list for each user based on various kinds of information. For quite a long time, most researchers have been pursing recommendation performances with predefined metrics, e.g., accuracy. However, in real-world applications, users select items from a huge item list by considering their internal personalized demand and external constraints. Thus, we argue that explicitly modeling the complex relations among items under domain-specific applications is an indispensable part for enhancing the recommendations. Actually, in this area, researchers have done some work to understand the item relations gradually from “implicit” to “explicit” views when recommending. To this end, in this paper, we conduct a survey of these recent advances on recommender systems from the perspective of the explicit item relation understanding. We organize these relevant studies from three types of item relations, i.e., combination-effect relations, sequence-dependence relations, and external-constraint relations. Specifically, the combination-effect relation and the sequence-dependence relation based work models the intra-group intrinsic relations of items from the user demand perspective, and the external-constraint relation emphasizes the external requirements for items. After that, we also propose our opinions on the open issues along the line of understanding item relations and suggest some future research directions in recommendation area.