Legal judgment prediction (LJP) is a crucial task in legal intelligence to predict charges, law articles and terms of penalties based on case fact description texts. Although existing methods perform well, they still have many shortcomings. First, the existing methods have significant limitations in understanding long documents, especially those based on RNNs and BERT. Secondly, the existing methods are not good at solving the problem of similar charges and do not fully and effectively integrate the information of law articles. To address the above problems, we propose a novel LJP method. Firstly, we improve the model’s comprehension of the whole document based on a graph neural network approach. Then, we design a graph attention network-based law article distinction extractor to distinguish similar law articles. Finally, we design a graph fusion method to fuse heterogeneous graphs of text and external knowledge (law article group distinction information). The experiments show that the method could effectively improve LJP performance. The experimental metrics are superior to the existing state of the art.

Legal judgment prediction (LJP) task aims to predict relevant law articles, charges and terms of penalty through the textual fact description of legal cases. However, existing methods face the core challenge is external legal knowledge under-exploitation, which hinders the performance of LJP. To overcome this limitation, a legal knowledge enhanced prompt learning based approach (LKEPL) is proposed. Our method first design the legal knowledge extraction which utilizes the model trained on our labeled legal dataset to identify legal entity relation knowledge and obtains law articles knowledge by the attention mechanism. Afterwards, our method employs legal knowledge projector to translate legal knowledge vectors into the transformer space. Then our method incorporates two kinds of knowledge by concatenating those vectors before the key and value matrix of each multi-head attention of BERT. Finally, our method designs the LJP task-specific prompt learning method. Experimental results on the CAIL2018 dataset demonstrate that our method outperforms existing strong baseline models. Furthermore, our method offers interpretability and shows effectiveness under few-shot scenarios.

Document-level relation extraction is a challenging task in information extraction, as it involves identifying semantic relations between entities that are dispersed throughout a document. Existing graph-based approaches often rely on simplistic methods to construct text graphs, which do not provide enough lexical and semantic information to accurately predict the relations between entity pairs. In this paper, we introduce a document-level relation extraction method called SKAMRR (Sememe Knowledge-enhanced Abstract Meaning Representation and Reasoning). First, we generate document-level abstract meaning representation graphs using rules and acquire entity nodes’ features through sufficient information propagation. Next, we construct inference graphs for entity pairs and utilize graph neural networks to obtain their representations for relation classification. Additionally, we propose the global adaptive loss to address the issue of long-tailed data. We conduct extensive experiments on four datasets DocRE, CDR, GDA, and HacRED. Our model achieves competitive results and its performance outperforms previous state-of-the-art methods on four datasets.

The charge prediction task aims to predict appropriate charges for a given legal case automatically, which still confronts some challenging problems such as performance improvement and confusing charges issue. In this paper, inspired by the impressive success of deep neural networks in legal intelligence field, we present an end-to-end framework named law article deduplication attention neural network, ADAN, to address these problems. The incorporation of hierarchical sequence encoder and attention mechanism is employed to learn better semantic representations of fact description texts. To distinguish confusing charges, we use the relevant law articles of a given case as auxiliary information, and propose a novel difference aggregation mechanism among similar law articles for extracting effective distinguishable features. The experimental results on real-world datasets show that the performance of our proposed model is significantly better than existing methods on all evaluation metrics.

As an essential branch of artificial intelligence, recommendation systems have gradually penetrated people’s daily lives. It is the active recommendation of goods or services of potential interest to users based on their preferences. Many recommendation methods have been proposed in both industry and academia. However, there are some limitations of previous recommendation methods: (1) Most of them do not consider the cross-correlation between data. (2) Many treat the recommendation process as a one-time act and do not consider the continuity of the recommendation system. To overcome these limitations, we propose a recommendation framework based on deep reinforcement learning techniques, known as DDRCN: a deep deterministic policy gradient recommendation framework incorporating deep cross networks. We use a Deep network and a Cross network to fit the cross relationships between the data, to obtain a representation of the user interaction data. The Actor-Critic network is designed to simulate the continuous interaction behavior of users through a greedy strategy. A deep deterministic policy gradient network is also used to train the recommendation model. Finally, we conduct experiments with two publicly available datasets and find that our proposed recommendation framework outperforms the baseline approach in the recall and ranking phases of recommendations.