毕业生知识库

With the development of information technology, various industries are constantly producing massive amounts of documents. It is an urgent need to efficiently and automatically process the complex information of these documents and to improve the efficiency of downstream applications. The document representation based on the graph neural network aims to model the topological information in documents using graph structures, which can obtain the encoded representation of documents with structured prior constraints, and provide more efficient support for downstream tasks. The main point of using graph neural networks to represent documents is how to identify the correlations between multi-granularity semantic nodes. The mainstream methods use the organizational structure or semantic structure of the document to construct this correlation, which is used to enhance the representation of the semantic information of the document itself. However, the representation enhanced based on the structure of the document itself is not optimized for the downstream task, and thus there will be deviations from the downstream targets, which can cause confusion in complex reasoning such as multi-document, multi-task, and multi-modal reasoning. To address this, based on the graph neural network, this thesis firstly studies the use of domain knowledge to achieve domain-specific representation enhancement of document representations. Based on the similar structure, this thesis then studies the estimation of causal associations between semantic nodes and downstream targets based on counterfactual reasoning. Subsequently, this thesis extends counterfacutal reasoning on the graph to an end-to-end form and uses its obtained causal associations to achieve document representation enhancement for downstream targets in complex scenarios of cross-document joint reasoning and cross-modality information alignment. The main research results of this thesis are summarized as follows:

1. Joint modeling of the document and label with clause interaction hypergraph

Enhancing the representation of documents with additional domain expert knowledge is an effective means to apply natural language processing techniques to vertical domains. Most of the existing methods inject external knowledge into the document representation process unidirectionally without constraining the representation of external knowledge, which leads to the lack of high-level semantic interaction with bidirectional feedback between the two. To this end, this thesis proposes an algorithm called joint modeling of the document and label with clause interaction hypergraph. In this algorithm, the thesis first proposes a clause interaction hypergraph to jointly model the document and the structured label set, which injects the structural description information of the label set into the document representation by the interaction connections, and uses the semantic structure of the document to constrain the representations of the label set based on the hypergraph. Subsequently, this thesis proposes a two-stage hybrid hypergraph convolution algorithm to iteratively represent the hypergraph structure and the simple graph structure in the clause interaction hypergraph with the graph neural network. Experiments based on the large-scale medical code assignment task show that the method proposed in this thesis can effectively inject domain knowledge into the document representation process, and successfully use the document semantic structure to impose semantic constraints on the representation of structured knowledge, which significantly improves the prediction precision of downstream tasks.

2. Counterfactual reasoning for document representation interpretability

Representing documents in high-risk scenarios such as healthcare and finance requires models that can provide effective interpretive support for the results. Most of the existing methods require the introduction of an additional expert knowledge base in the model interpretation process. However, in many subdivision scenarios, it is difficult to obtain sufficiently accurate knowledge annotation due to human limitations. To this end, this thesis proposes a cascaded counterfactual reasoning document representation interpretation algorithm based on graph neural networks. The algorithm discretizes unstructured documents through the graph structure, and then constructs a hierarchical graph containing different granularity semantic units, whose representations are used for the prediction of downstream tasks. Based on the completed trained document representation model, this thesis proposes two counterfactual reasoning methods based on node treatment and edge treatment respectively to estimate the association between different semantic units and downstream task targets, and extract the causal graph structure oriented to the downstream task targets. Strongly associated semantic units are utilized to provide interpretability of the model for each instance. Experiments based on real electronic medical records show that the algorithm proposed in this thesis can effectively realize automatic diagnosis based on electronic medical records and provide interpretable support in accordance with medical logic for the diagnosis.

3. Local causal association graph for document confusing information filtering

It is a long-standing challenge to perform joint reasoning on a large number of documents retrieved from the open domain to obtain the desired target information. The difficulty lies in the fact that there is a lot of confusing information irrelevant to downstream tasks in the retrieved documents, and they will become noises that cause serious interference in global reasoning. To this end, this thesis proposes a local causal association graph model for confusing information filtering in the document. In this model, this thesis first improves the cascaded counterfactual reasoning algorithm of work 2 to an end-to-end form, which can estimate the causal effects of specific semantic units in the process of joint optimization with downstream tasks. Subsequently, based on the document structure information, this thesis proposes a local causal association graph to model the multi-document joint reasoning process, which uses the causal effect of semantic unit nodes to filter confusing information nodes, so as to construct the optimal causal graph structure for downstream tasks and enhance the semantic representation of valid information nodes. Experiments in the cross-document relation extraction task show that the proposed method in this thesis can accurately distinguish valid and confusing semantic units in complex information and filter the confusing information in global reasoning, which significantly improves the accuracy of downstream tasks. In addition, this thesis also verifies the mutual enhancement between the accuracy of causal reasoning and the ability of document representation based on the graph neural network.

4. Feature attribution graph for causal alignment of document cross-modality information

Incorporating cross-modality information in the process of document representation is a trend that pushes natural language processing technology into the general domain. The difficulty is how to construct the alignment pattern between cross-modality information. Constructing associations between cross-modality features through semantic similarity matching is the current mainstream approach. However, this correlation is not directly related to the downstream task, which leads to bias between the cross-modality representations constructed based on semantic similarity matches and downstream tasks. To this end, this thesis proposes a feature attribution graph model for the causal alignment of cross-modality information. In this model, this thesis first proposes a causal reasoning based feature attribution algorithm for multi-task scenarios, which is able to evaluate the importance of fine-grained features relative to a specific downstream task. Subsequently, this thesis constructs a feature attribution graph for cross-modality information alignment, which uses feature attribution values to calculate alignment weights between cross-modality node features to achieve cross-modality document representation for a specific downstream task. Experiments in a multimodal desire understanding task show that the proposed method in this thesis can effectively reduce the bias between cross-modality document representation and downstream tasks, and obtain significant performance improvements on multiple subtasks.