毕业生知识库

1.Document-level entity relation extraction based on co-reference resolution: Entities in document text usually contain multiple entity mentions, which can be divided into noun phrase mentions and pronoun mentions according to the type of mention. Existing methods only use noun phrase mentions and ignore pronoun mentions in entity representation. Pronoun mentions account for a large proportion of entity references, and using them to represent entities enriches entity representations while also shortening the distance between entities, thereby reducing the difficulty of entity relationship reasoning and improving the model’s inference performance. In response to this issue, this article proposes a document level entity relationship extraction method based on referential resolution. The pointer generation network model is used to align the noun phrase mentions and pronoun mentions in the text, and then the aligned pronoun and noun phrase mentions are used together as entity representations to construct a text graph to infer entity relationships. This method achieved an F1 value of 62.46% on the DocRED datasets, exceeding the benchmark model by 1.42%, achieving the optimal effect and proving its effectiveness.

2.Document-level entity relation extraction based on information bottleneck: In the task of document-level entity relationship extraction, there are multiple sentences, entities, and mentions in the text. In the process of identifying entity relationships, different levels of information need to be combined for logical reasoning. The existing methods use all textual information as the basis for inferring entity relationships, which can generate certain noise due to data redundancy and interfere with the effectiveness of entity relationship reasoning. In response to this issue, this article proposes a document-level entity relationship extraction method based on information bottleneck, which uses entity mentions and sentences as nodes to construct a text graph. During the update process of node information in the graph, a graph random attention method based on information bottleneck is used to automatically focus on sentences and mentions nodes related to entity relationship discrimination, thereby reducing the negative impact of irrelevant information on entity relationship discrimination, To achieve the goal of improving model extraction efficiency. Compared to the graph attention method using a complete graph, this method improves the F1 value by 1.85% on the DocRED datasets, proving the effectiveness of this method.