Font Size: a A A

Multi-level And Multi-modal Named Entity Recognition

Posted on:2022-07-03Degree:MasterType:Thesis
Country:ChinaCandidate:C M ZhengFull Text:PDF
GTID:2518306569481514Subject:Software engineering
Abstract/Summary:PDF Full Text Request
Knowledge graph describes the concept,entity and its relationship in the objective world in a structured form,which is a way to organize,manage and understand the mass information data.Named entity recognition(NER)is a key step in building a knowledge graph.The NER task needs to locate named entities from unstructured text and classify them into specific categories,such as names,locations,and organizations.Although the existing neural network-based named entity model has achieved great success in some fields,there are two obvious shortcomings: First,the existing methods mainly focus on the task of non-nested named entity recognition,and ignore the entity's problem of multi-level nesting;second,the existing methods mainly focus on the text in the news media field,and the performance of the short text in the social media field is drastically reduced due to insufficient context information.To Address the problem of multi-level nesting of entities,we propose a multi-level nested named entity recognition model based on boundary information.Nested entities are common in many areas.Because words in nested entities often contain multiple tags,the previous single-level sequence tagging model cannot handle the nesting problem.The multi-layer sequence annotation model has serious layer-to-layer error propagation.The exhaustive fragment model often has a lot of boundary errors due to the lack of guidance of boundary tags.We propose a neural network model based on boundary perception.Through boundary detection,the model can generate boundary-related candidate entity regions.The text representations of these regions are used to predict nested entity categories.Compared with the existing multi-layer sequence labeling model,the neural network model based on boundary perception requires smaller computing resources and can slow down the error propagation in the multi-layer model,thereby achieving better prediction results.To address the entity recognition problem of social media short text,we respectively proposed a multi-modal named entity recognition model based on adversarial training and bilinear attention mechanism.Visual context information can usually assist in more accurate identification of short text social media entities.In the past,multi-modal named entity recognition work only used the whole image or label information,and ignored the correspondence between fine-grained visual objects and named entities.At the same time,due to the spatial difference between visual representation and textual representation,simple splicing of the two representations will bring a semantic shift.In response to the above problems,we proposed a multi-modal named entity recognition model based on adversarial training and bilinear attention mechanism.The model can extract entity-related features from visual objects and text at the same time,and use adversarial training to map two different modal representations to the same semantic vector space.Therefore,the semantic information contained in the image can be transferred and used to assist in accurately identifying the named entity in the text.Our model outperformed all previous state-of-the-art models on the Twitter multimodal data set in the social media field.At the same time,we have proved the model's effectiveness through data distribution visualization,attention weight visualization,and substantial case analysis.
Keywords/Search Tags:Multi-level Nested Named Entity Recognition, Multi-modal Named Entity Recognition, Knowledge Graph, Social Media Short Text
PDF Full Text Request
Related items