Offline meta reinforcement learning (OMRL) has emerged as a promising approach for interaction avoidance and strong generalization performance by leveraging pre-collected data and meta-learning techniques. Previous context-based approaches predominantly rely on the intuition that alternating optimization between the context encoder and the policy can lead to performance improvements, as long as the context encoder follows the principle of maximizing the mutual information between the task variable $M$ and its latent representation $Z$ ($I(Z;M)$) while the policy adopts the standard offline reinforcement learning (RL) algorithms conditioning on the learned task representation. Despite promising results, the theoretical justification of performance improvements for such intuition remains underexplored. Inspired by the return discrepancy scheme in the model-based RL field, we find that the previous optimization framework can be linked with the general RL objective of maximizing the expected return, thereby explaining performance improvements. Furthermore, after scrutinizing this optimization framework, we observe that the condition for monotonic performance improvements does not consider the variation of the task representation. When these variations are considered, the previously established condition may no longer be sufficient to ensure monotonicity, thereby impairing the optimization process. We name this issue \underline{task representation shift} and theoretically prove that the monotonic performance improvements can be guaranteed with appropriate context encoder updates. We use different settings to rein in the task representation shift on three widely adopted training objectives concerning maximizing $I(Z;M)$ across different data qualities. Empirical results show that reining in the task representation shift can indeed improve performance. Our work opens up a new avenue for OMRL, leading to a better understanding between the task representation and performance improvements.

Nowadays, more and more researchers are committed to applying artificial intelligence technology to the legal field to support decision-making. Charge prediction is a subtask of legal judgment prediction (LJP). Its purpose is to analyze the fact description of natural language text and predict a charge corresponding to a case. At present, most of the research takes charge prediction as a multiclass classification task, which leads to unsatisfactory interpretation due to the weak semantic correlation between the fact descriptions and charge labels. In order to solve this problem, we propose a method of generative evidence mining based on prompt learning. Specifically, in the training phase, we reformulate the charge labels into the prompt template that we design to enhance the semantic correlation between the charge labels and the fact descriptions. In the testing phase, the charge labels are generated via the model based on prompt learning. Meanwhile, we calculate the attention score of each sentence from the multihead self-attention in the transformer encoder and choose the sentence with the highest attention score as the evidence. Our experimental results on a real dataset show that our method is better than the traditional fine-tuning-based classification method.

Exploring the high-effect factors of residents’ happiness is good for a wide range of policy-making for economics and politics in most countries. Limited to the concerns of sociologists, previous efforts manually predefined the relationship among multifactors to analyze the factor–factor interactions with high interpretability by regression models. Recently, deep learning methods show great promising prediction accuracy by automatically learning additive interaction between factors, while it meets the challenges of interpretability. To this end, an unbiased post-hoc and model-agnostic method is promising to quantitatively interpret the results of the happiness prediction model. Thus, this article proposes a novel solution based on the deep neural network (DNN) and the Shapley value to compute the factor–factor interactions in different coalitions based on coalitional game theory. Aiming at evaluating the pairwise interactions interpretability quality of our solution, experiments are conducted on the Chinese General Social Survey (CGSS) and European Social Survey (ESS) questionnaire datasets. By systematic reviews, the experimental results are highly consistent with academic studies in social science. Analyzed by different factor categories, the new finding is that some factors, e.g., body mass index (BMI) and social media, play a crucial role in happiness prediction but are rarely considered in social science studies. Specifically, there are some heavy interactions across multicategories, such as personal information (health and age) with economics (insurance and income). Therefore, this solution can theoretically support the implications of social decision-making.