Zhaorui Jian

In the field of Chinese-named entity recognition, recent research has sparked new interest by combining lexical features with character-based methods. Although this vocabulary enhancement method provides a new perspective, it faces two main challenges: firstly, using character-by-character matching can easily lead to conflicts during the vocabulary matching process. Although existing solutions attempt to alleviate this problem by obtaining semantic information about words, they still lack sufficient temporal sequential or global information acquisition; secondly, due to the limitations of dictionaries, there may be words in a sentence that do not match the dictionary. In this situation, existing vocabulary enhancement methods cannot effectively play a role. To address these issues, this paper proposes a method based on lexical information and spatial features. This method carefully considers the neighborhood and overlap relationships of characters in vocabulary and establishes global bidirectional semantic and temporal sequential information to effectively address the impact of conflicting vocabulary and character fusion on entity segmentation. Secondly, the attention score matrix extracted by the point-by-point convolutional network captures the local spatial relationship between characters without fused vocabulary information and characters with fused vocabulary information, aiming to compensate for information loss and strengthen spatial connections. The comparison results with the baseline model show that the SISF method proposed in this paper improves the F1 metric by 0.72%, 3.12%, 1.07%, and 0.37% on the Resume, Weibo, Ontonotes 4.0, and MSRA datasets, respectively.

Distantly supervised relation extraction (DSRE) utilizes an external knowledge base to automatically label a corpus, which inevitably leads to the problem of mislabeling. Existing approaches utilize BERT to provide instances and relation embeddings to capture a wide set of relations and address the noise problem. However, the method suffers from a single method of textual information processing, underutilizing the feature information of entity pairs in the relation embeddings part and being interfered with by noisy labels when classifying multiple labels. For this reason, we propose the contextual information interaction and relation embeddings (CIRE) method. First, we utilize BERT and Bi-LSTM to construct a neural network model to enhance contextual information interaction by filtering and supplementing sequence information through the error repair capability of the Bi-LSTM gating mechanism. At the same time, we combine the vector difference between entity pairs and entity pairs in the relation embeddings layer to improve the relation embeddings accuracy. Finally, we choose sparse softmax as the classifier, which improves the ability to control the noise categories by controlling the number of output categories. The experimental results show that our method significantly outperforms the baseline method and improves the AUC metric by 2.6% on the NYT2010 dataset.

Transforming the task of information extraction into a machine reading comprehension (MRC) framework has shown promising results. The MRC model takes the context and query as the inputs to the encoder, and the decoder extracts one or more text spans as answers (entities and relationships) from the text. Existing approaches typically use multi-layer encoders, such as Transformers, to generate hidden features of the source sequence. However, increasing the number of encoder layers can lead to the granularity of the representation becoming coarser and the hidden features of different words becoming more similar, potentially leading to the model’s misjudgment. To address this issue, a new method called the multi-granularity attention multi-scale self-learning network (MAML-NET) is proposed, which enhances the model’s understanding ability by utilizing different granularity representations of the source sequence. Additionally, MAML-NET can independently learn task-related information from both global and local dimensions based on the learned multi-granularity features through the proposed multi-scale self-learning attention mechanism. The experimental results on two information extraction tasks, named entity recognition and entity relationship extraction, demonstrated that the method was superior to the method based on machine reading comprehension and achieved the best performance on the five benchmark tests.