Search

Recognizing entities and extracting their relations are significant tasks for information extraction. Many recent works focus on leveraging entity markers for better span (pair) representations, which have achieved promising performance. Nevertheless, existing works have two shortcomings: (1) previous markers are randomly embedded into distributed representations, which ignores semantic information relevant to the targeted tokens that we want to highlight. (2) most works simply implant markers into a sentence, which is weak to encode the interrelation between multiple tokens. In this work, we propose a marker collaborating model for entity and relation extraction, which is composed of two modules. The first module is semantic marker module, where two kinds of semantic markers are proposed to leverage semantic features from the PLMs. The second module is collaborative marker attention module, where a marker-oriented directional attention is proposed to model the interaction between multiple tokens. Our model yields SOTA performance on three standard benchmarks (ACE04, ACE05 and SciERC).

The court system mainly has two modes: manual designated division and simple random division.The above method cannot achieve automatic matching of persons and cases, and there are drawbacks such as money cases and relationship cases.At present, the research on division method mainly has two difficulties: judge’s representation and case matching.Combining the judge’s historical trial data, the judge’s expertise in the judge’s representation was integrated, and a judge representation method that integrates the quality of the trial was proposed.Then, the abstract semantic feature vectors of different granularities in the case representation and the judge representation were learned through the convolutional neural network, the cosine similarity between the case and the feature vectors of multiple judges was calculated, and vector similarity was used to indicate the matching degree between the case and the judge, the top N judges with high matching value were output as recommended judges for the case.Experiments with real data from a court in Guizhou Province, and the results show that the accuracy of the method for recommending judges is 80% higher than the traditional method.

Extraction of legal elements is an important basis for judicial intelligent auxiliary applications, and its purpose is to identify the key elements involved in the judgment document.In the past, extracting legal elements usually used multi-label classification methods for modeling.These methods mainly relied on the text features of the judgment document, thereby ignoring the label features.Besides, due to the imbalanced data problem in judicial data sets, the classification method will lead to poor model performance because of too many negative examples.To solve the above problems, a legal element extraction method based on BERT reading comprehension framework was proposed.This method constructed auxiliary questions with label information and legal prior knowledge, and used the machine reading comprehension model based on BERT to establish the semantic associations between question and judgment document.And this method added special tokens before and after the label in the question to enhance the learning ability of the model.Experiments were conducted on the legal element extraction data sets of the CAIL2019.Experiment results show that the performance is improved significantly, and the F1 value has been increased by 2.7%, 11.3%, and 5.6% respectively on the data sets of marriage and family case, labor dispute case, and loan contract dispute case.

Charge prediction is to find the appropriate charges based on the facts of the given case.Existing methods mainly use text content for classification, but they cannot effectively use case elements.For the shortcomings of the existing methods, the method of accusation prediction based on the sequence of case elements was put forward.The way expressed the case factual processes as a series of case elements with “behavior” as the core and time-series relationship.Then graph convolutional network was used to represent.Finally, the semantic features of the text were fused to predict the crime.Experiments show that this method has better prediction performance than existing methods.Meanwhile, this method also has a good performance for the distinction between easily confusing charges.

Aiming at the problem of difficult recognition and uniqueness of Chinese predicate head, a Highway-BiLSTM model was proposed.Firstly, multi-layer BiLSTM networks were used to capture multi-granular semantic dependence in a sentence.Then, a Highway network was adopted to alleviate the problem of gradient disappearance.Finally, the output path was optimized by a constraint layer which was designed to guarantee the uniqueness of predicate head.The experimental results show that the proposed method effectively improves the performance of predicate head recognition.

The assignment of cases means the court assigns cases to specific judges. The traditional case assignment methods, based on the facts of a case, are weak in the analysis of semantic structure of the case not considering the judges' expertise. By analyzing judges' trial logic, we find that the order of criminal behaviors affects the final judgement. To solve these problems, we regard intelligent case assignment as a text-matching problem, and propose an intelligent case assignment method based on the chain of criminal behavior elements. This method introduces the chain of criminal behavior elements to enhance the structured semantic analysis of the case. We build a BCTA (Bert-Cnn-Transformer-Attention) model to achieve intelligent case assignment. This model integrates a judge's expertise in the judge's presentation, thus recommending the most compatible judge for the case. Comparing the traditional case assignment methods, our BCTA model obtains 84% absolutely considerable improvement under P@1. In addition, comparing other classic text matching models, our BCTA model achieves an absolute considerable improvement of 4% under P@1 and 9% under Macro F1. Experiments conducted on real-world data set demonstrate the superiority of our method.

In response to the problems traditional multi-view document clustering methods separate the multi-view document representation from the clustering process and ignore the complementary characteristics of multi-view document clustering,an iterative algorithm for complementary multi-view document clustering——CMDC was proposed,in which the multi-view document clustering process and the multi-view feature adjustment were conducted in a mutually unified manner.In CMDC algorithm,complementary text documents were selected from the clustering results to aid adjusting the contribution of view features via learning a local measurement metric of each document view.The complementary text document of the results among the dimensionality clusters was selected by CMDC,and used to promote the feature tuning of the clusters.The partition consistency of the multi-dimensional document clustering was solved by the measure consistency of the dimensions.Experimental results show that CMDC effectively improves multi-dimensional clustering performance.

The task of relation recognition identifies semantic relationships between two named entities in a sentence. In neural network based models, a convolutional layer is often conducted to extract representative local features of a sentence. The convolution operation is implemented through a whole sentence, without considering the structure of a sentence. Because the task to recognize entity relation is processed in sentence level, many ambiguous phenomena (e.g., polysemy) are influential rather than in a document. Capturing structural information of a sentence is helpful to solve this problem. In this paper, a multi-channel framework is presented, which uses two named entities to divide a sentence into several channels. Each channel is stacked with layered neural networks. These channels do not interact during recurrent propagation, which enables a neural network to learn different representations. In our experiments, it outperforms the widely used position embedding approach. Comparing with the state-of-the-art approaches, its performance shows a meaningful improvement.

Biomedical named entity recognition (BNER) is an important task in biomedical natural language processing, in which neologisms (new terms, words) are coined constantly. Most of the existing work can only identify biomedical named entities with flattened structures and ignore nested biomedical named entities and discontinuous biomedical named entities. Because biomedical domains often use nested structures to represent semantic information of named entities, existing methods fail to utilize abundant information when processing biomedical texts. This paper focuses on identifying nested biomedical named entities using a boundary assembly (BA) model, which is a cascading framework consisting of three steps. First, start and end named entity boundaries are identified and then assembled into named entity candidates. Finally, a classifier is implemented for filtering false named entities. Our approach is effective in handling nesting and discontinuous problems in biomedical named entity recognition tasks. It improves the performance considerably, achieving an F1-score of 81.34% on the GENIA dataset.

There is large and growing amounts of textual data that contains information about human activities. Mining interesting knowledge from this textual data is a challenging task because it consists of unstructured or semistructured text that are written in natural language. In the field of artificial intelligence, event-oriented techniques are helpful in addressing this problem, where information retrieval (IR), information extraction (IE) and graph methods (GMs) are three of the most important paradigms in supporting event-oriented processing. In recent years, due to information explosions, textual event detection and recognition have received extensive research attention and achieved great success. Many surveys have been conducted to retrospectively assess the development of event detection. However, until now, all of these surveys have focused on only a single aspect of IR, IE or GMs. There is no research that provides a complete introduction or a comparison of IR, IE, and GMs. In this article, a survey about these techniques is provided from a broader perspective, and a convenient and comprehensive comparison of these techniques is given. The hallmark of this article is that it is the first survey that combines IR, IE and GMs in a single frame and will therefore benefit researchers by acting as a reference in this field.

Mining a document structure from multiple data sources in terms of their underlying topics has become an important task of document clustering. The traditional document clustering approach cannot be applied directly to the multi-source document clustering problem. There are three typical difficulties: 1) The topics of different data sources are related but not the same. 2) Usually, each data source has its own focus on topics. 3) The number of clusters of the data sources are not necessarily the same and are not known beforehand. In this paper, based on our previous research, we design a novel multi-source document clustering model, namely, the hierarchical Dirichlet multinomial allocation (HDMA) model, to solve all the above problems. The HDMA model is investigated with a two-step hierarchical topic generation process. Topics are learnt to share their general characteristics across data source, while at the same time preserve the local characteristics of the data source. Each data source is applied with an exclusive topic partition to learn the source-level topic emphasis. A Gibbs sampling algorithm is then used to learn the number of clusters for each data source as well as the parameters of the HDMA model at the same time. Experimental results demonstrate that the HDMA model is effective.

The context of a sentence is composed of a limited number of words. This leads to the feature sparsity problem whereby the sentence's meaning is easily influenced by language phenomena such as polysemy, ambiguity and puns. To resolve these problems, the set space model (SSM) uses language characteristics to group features of a sentence into different sets. Afterwards, the proposed feature calculus is used to capture the structural information of the sentence. Experiments have shown that this approach to the relation recognition task is effective. However, at least three weaknesses remain. First, due to the lack of a probabilistic explanation, several aspects of SSM (e.g., filter selection) have not yet been covered. Second, the existing studies have only provided an outline of SSM, and many issues remain unclear. To understand this approach, it is necessary to discuss a suitable example in detail. Third, SSM has been applied only to the task of relation recognition. Case studies of more typical topics (e.g., named entity recognition) will help illustrate the use of SSM's methodology to manipulate features. This paper develops SSM to cover these problems. It describes a systematic and novel approach to manipulating features of a sentence. In the experimental part, two typical information extraction tasks are performed to demonstrate SSM's capabilities. Two case studies are considered, and favorable improvements are observed. All of the obtained results surpass those of compared approaches. The experiments also show the influence of sentence structural information on information extraction.

Sentencing prediction is an important direction of artificial intelligence applied to the judicial field. The purpose is to predict the trial sentence for the case based on the description of the case in the adjudication documents. Traditional methods mainly use neural networks exclusively, which are trained on a large amount of data to encode textual information and then directly regress or classify out the sentence. This shows that machine learning methods are effective, but are extremely dependent on the amount of data. We found that there is still external knowledge such as laws and regulations that are not used. Moreover, the prediction of sentences in these methods does not fit well with the trial process. Thus, we propose a sentence prediction method that incorporates trial logic based on abductive learning, called SPITL. The logic of the trial is reflected in two aspects: one is that the process of sentence prediction is more in line with the logic of the trial, and the other is that external knowledge, such as legal texts, is utilized in the process of sentence prediction. Specifically, we establish a legal knowledge base for the characteristics of theft cases, translating relevant laws and legal interpretations into first-order logic. At the same time, we designed the process of sentence prediction according to the trial process by dividing it into key circumstance element identification and sentence calculation. We fused the legal knowledge base as weakly supervised information into a neural network through the combination of logical inference and machine learning. Furthermore, a sentencing calculation method that is more consistent with the sentencing rules is proposed with reference to the Sentencing Guidelines. Under the condition of the same training data, the effect of this model in the experiment of responding to the legal documents of theft cases was improved compared with state-of-the-art models without domain knowledge. The results are not only more accurate as a sentencing aid in the judicial trial process, but also more explanatory.

The task to extract relations tries to identify relationships between two named entities in a sentence. Because a sentence usually contains several named entities, capturing structural information of a sentence is important to support this task. Currently, graph neural networks are widely implemented to support relation extraction, in which dependency trees are employed to generate adjacent matrices for encoding structural information of a sentence. Because parsing a sentence is error-prone, it influences the performance of a graph neural network. On the other hand, a sentence is structuralized by several named entities, which precisely segment a sentence into several parts. Different features can be combined by prior knowledge and experience, which are effective to initialize a symmetric adjacent matrix for a graph neural network. Based on this phenomenon, we proposed a feature combination-based graph convolutional neural network model (FC-GCN). It has the advantages of encoding structural information of a sentence, considering prior knowledge, and avoiding errors caused by parsing. In the experiments, the results show significant improvement, which outperform existing state-of-the-art performances.

Relation extraction aims to extract semantic relationships between two specified named entities in a sentence. Because a sentence often contains several named entity pairs, a neural network is easily bewildered when learning a relation representation without position and semantic information about the considered entity pair. In this paper, instead of learning an abstract representation from raw inputs, task-related entity indicators are designed to enable a deep neural network to concentrate on the task-relevant information. By implanting entity indicators into a relation instance, the neural network is effective for encoding syntactic and semantic information about a relation instance. Organized, structured and unified entity indicators can make the similarity between sentences that possess the same or similar entity pair and the internal symmetry of one sentence more obviously. In the experiment, a systemic analysis was conducted to evaluate the impact of entity indicators on relation extraction. This method has achieved state-of-the-art performance, exceeding the compared methods by more than 3.7%, 5.0% and 11.2% in F1 score on the ACE Chinese corpus, ACE English corpus and Chinese literature text corpus, respectively.

In view of the common problems and practical needs in the construction of “Intelligent Court”,the framework of “Intelligent Court” data fusion analysis and integration application demonstration platform was introduced.The research ideas and implementation paths of common key technologies in the construction of “Intelligent Court” were put forward from four aspects:deep semantic learning of judicial big data,judicial data fusion based on knowledge map,judicial data security protection and privacy protection,and visualization of judicial data fusion analysis.Finally,taking evidence extraction,criminal chain construction and legal provisions recommendation as examples,the application effect of data fusion analysis and integrated application demonstration platform was analyzed.The research results have certain reference value for realizing the goal of building a new generation of “Intelligent Court” with the judicial data of courts as the core.

Legal judgment document summarization, as a task specific to the legal domain, involves automatically generating a fluent, informative, and well-organized summary from the original legal judgment document. Unlike traditional text summarization tasks, this domain-specific task places higher demands on content accuracy and completeness in the summary, while also requiring the preservation of the professional expression found in the original text. Consequently, conventional summarization methods often struggle to perform effectively in the legal domain. In response to this challenge, this paper introduces a hybrid summarization model tailored for legal judgment documents. Our model harnesses the strengths of both extractive and abstractive summarization methods, incorporating domain knowledge to enhance the summary generation process. We conduct extensive experiments to verify the effectiveness of our proposed method and compare the results with a baseline using ROUGE evaluation metrics. The experimental findings highlight that our model excels in providing more accurate and readable summarizations compared to traditional methods. [ABSTRACT FROM AUTHOR]