To address the issue of insufficient extraction of semantic feature information with different scales and the lack of focus on crucial information when obtaining sentence semantic information by Convolutional Neural Network （CNN）-based relation extraction， a model for relation extraction based on a multi-scale hybrid attention CNN was proposed. Firstly， relation extraction was modeled as label prediction with two-dimensional representation. Secondly， by extracting and fusing multi-scale feature information， finer-grained multi-scale spatial information was obtained. Thirdly， through the combination of attention and convolution， the feature maps were refined adaptively to make the model concentrate on important contextual information. Finally， two predictors were used jointly to predict the relation labels between entity pairs. Experimental results demonstrate that the multi-scale hybrid convolutional attention model can capture multi-scale semantic feature information，And the key information in channels and spatial locations was captured by the channel attention and spatial attention by assigning appropriate weights， thereby improving the performance of relation extraction. The proposed model achieves F1 scores of 90.32% on SemEval （SemEval-2010 task 8） dataset， 70.74% on TACRED （TAC Relation Extraction Dataset）， 85.71% on Re-TACRED （Revised-TACRED）， and 89.66% on SciERC （Entities， Relations， and Coreference for Scientific knowledge graph construction）.

Existing methods for answer acquisition based on pre-trained language models may suffer from inaccuracies in predicting boundaries， a boundary-aware approach for span-based extraction Machine Reading Comprehension （MRC） is proposed to mitigate this issue. Firstly， special characters were introduced to mark the question boundary during the question input stage， enhancing the semantic information of the question to improve boundary perception. Secondly， during the answer prediction stage， an answer boundary regressor was constructed to facilitate semantic interaction between the perceived question boundary and the output of the predicted answer boundary. Lastly， the biased predicted answer boundary was further adjusted based on the post-interaction semantic information to calibrate the predicted answers. Experimental results demonstrate that when compared to the SpanBERT （Span-based Bidirectional Encoder Representation from Transformers）， the proposed method improves the F1 value by 0.2 percentage points and the Exact Match （EM） value by 0.9 percentage points on the public dataset SQuAD （Stanford Question Answering Dataset）1.1， it achieved improvements of 0.7 percentage points in both F1 score and EM value on the HotpotQA （Hotpot Question Answering） dataset， and it improved the F1 score by 2.8 percentage points and the EM value by 3.3 percentage points on the NewsQA （News Question Answering） dataset. The effectiveness of this method is rooted in its capacity to enhance the model’s perception of question boundary information and to accomplish the calibration of predicted answer boundary. Consequently， it results in an enhancement of system accuracy in applications such as intelligent question answering and intelligent customer service when dealing with text data comprehension and analysis.

Aiming at the problem that the existing deep clustering methods can not efficiently divide event types without considering event information and its structural characteristics， a Deep Event Clustering method based on Event Representation and Contrastive Learning （DEC_ERCL） was proposed. Firstly， information recognition was utilized to identify structured event information from unstructured text， thus the impact of redundant information on event semantics was avoided. Secondly， the structural information of the event was integrated into the autoencoder to learn the low-dimensional dense event representation， which was used as the basis for downstream clustering. Finally， in order to effectively model the subtle differences between events， a contrast loss with multiple positive examples was added to the feature learning process. Experimental results on the datasets DuEE， FewFC， Military and ACE2005 show that the proposed method performs better than other deep clustering methods in accuracy and Normalized Mutual Information （NMI） evaluation indexes. Compared with the suboptimal method， the accuracy of DEC_ERCL is increased by 17.85%，9.26%，7.36% and 33.54%， respectively， indicating that DEC_ERCL has better event clustering effect.

To tackle the difficulty in semantic mining of entity relations and biased relation prediction in Relation Extraction （RE） tasks， a RE method based on Mask prompt and Gated Memory Network Calibration （MGMNC） was proposed. First， the latent semantics between entities within the Pre-trained Language Model （PLM） semantic space was learned through the utilization of masks in prompts. By constructing a mask attention weight matrix， the discrete masked semantic spaces were interconnected. Then， the gated calibration networks were used to integrate the masked representations containing entity and relation semantics into the global semantics of the sentence. Besides， these calibrated representations were served as prompts to adjust the relation information， and the final representation of the calibrated sentence was mapped to the corresponding relation class. Finally， the potential of PLM was fully exploited by the proposed approach through harnessing masks in prompts and combining them with the advantages of traditional fine-tuning methods. The experimental results highlight the effectiveness of the proposed method. On the SemEval （SemEval-2010 Task 8） dataset， the F1 score reached impressive 91.4%， outperforming the RELA （Relation Extraction with Label Augmentation） generative method by 1.0 percentage point. Additionally， the F1 scores on the SciERC （Entities， Relations， and Coreference for Scientific knowledge graph construction） and CLTC （Chinese Literature Text Corpus） datasets were remarkable， achieving 91.0% and 82.8% respectively. The effectiveness of the proposed method was evident as it consistently outperformed the comparative methods on all three datasets mentioned above. Furthermore， the proposed method achieved superior extraction performance compared to generative methods.

With the application of artificial intelligence technology in the judicial field， charge prediction based on case description has become an important research content. It aims at predicting the charges according to the case description. The terms of case contents are professional， and the description is concise and rigorous. However， the existing methods often rely on text features， but ignore the difference of relevant elements and lack effective utilization of elements of action words in diverse cases. To solve the above problems， a multi-task learning model of charge prediction based on action words was proposed. Firstly， the spans of action words were generated by boundary identifier， and then the core contents of the case were extracted. Secondly， the subordinate charge was predicted by constructing the structure features of action words. Finally， identification of action words and charge prediction were uniformly modeled， which enhanced the generalization of the model by sharing parameters. A multi-task dataset with action word identification and charge prediction was constructed for model verification. The experimental results show that the proposed model achieves the F value of 83.27% for action word identification task， and the F value of 84.29% for charge prediction task； compared with BERT-CNN， the F value respectively increases by 0.57% and 2.61%， which verifies the advantage of the proposed model in identification of action words and charge prediction.

Aiming at the problem of poor recognition of sentencing circumstances in adjudication documents caused by the lack of labeled data， low quality of labeling and existence of strong logicality in judicial field， a sentencing circumstance recognition model based on abductive learning named ABL-CON （ABductive Learning in CONfidence） was proposed. Firstly， combining with neural network and domain logic inference， through the semi-supervised method， a confidence learning method was used to characterize the confidence of circumstance recognition. Then， the illogical error circumstances generated by neural network of the unlabeled data were corrected， and the recognition model was retrained to improve the recognition accuracy. Experimental results on the self-constructed judicial dataset show that the ABL-CON model using 50% labeled data and 50% unlabeled data achieves 90.35% and 90.58% in Macro_F1 and Micro_F1， respectively， which is better than BERT （Bidirectional Encoder Representations from Transformers） and SS-ABL （Semi-Supervised ABductive Learning） under the same conditions， and also surpasses the BERT model using 100% labeled data. The ABL-CON model can effectively improve the logical rationality of labels as well as the recognition ability of labels by correcting illogical labels through logical abductive correctness.

Relation extraction aims to extract the semantic relationships between entities from the text. As the upper-level task of relation extraction， entity recognition will generate errors and spread them to relation extraction， resulting in cascading errors. Compared with entities， entity boundaries have small granularity and ambiguity， making them easier to recognize. Therefore， a relationship extraction method based on entity boundary combination was proposed to realize relation extraction by skipping the entity and combining the entity boundaries in pairs. Since the boundary performance is higher than the entity performance， the problem of error propagation was alleviated； in addition， the performance was further improved by adding the type features and location features of entities through the feature combination method， which reduced the impact caused by error propagation. Experimental results on ACE 2005 English dataset show that the proposed method outperforms the table-sequence encoders method by 8.61 percentage points on Macro average F1-score.