Linear late-fusion semantic structural retrieval
Systems and methods for generating a fusion score between electronic documents. The method includes receiving a first electronic document by a document management system. The method further includes extracting a first set of features from the first electronic document including at least one feature type indicating the hierarchical structure of the first electronic document. The method also includes receiving a second electronic document by the document management server. The method further includes extracting a second set of features from the second electronic document including at least one feature type indicating the hierarchical structure of the second electronic document. The method further includes generating a fusion score based on a comparison of the first set of features and the second set of features.
1. A method comprising:
receiving a first electronic document by a document management server comprising a computer processor;
extracting a first set of features from the first electronic document, the first set of features comprising a hierarchical structure of the first electronic document and a content of the first electronic document;
receiving a second electronic document by the document management server;
extracting, by the document management server, a second set of features from the second electronic document, the second set of features comprising a hierarchical structure of the second electronic document and a content of the second electronic document;
generating, by the document management server, a similarity score for each feature by comparing each feature in the first set of features to a corresponding feature in the second set of features;
generating a weighted feature, by a machine learning model, for each feature of the first electronic document and the second electronic document, the weighted feature based on feedback from a previous similarity score; and
generating, by the document management server, a fusion score comprising a weighted combination of the similarity scores, wherein the weighted combination is based on the weighted feature, and the fusion score represents a similarity between the first electronic document and the second electronic document.
2. The method of claim 1 , wherein the step of extracting the first set of features comprises extracting a tree path structure from the first electronic document, wherein the tree path structure comprises a plurality of nodes.
3. The method of claim 2 , wherein the tree path structure further comprises semantic data for each node of the plurality of nodes, wherein the semantic data identifies a position of each node within the tree path structure, and wherein the semantic data includes structural and contextual information corresponding to each node.
4. The method of claim 1 , wherein generating the similarity score for each feature further comprises computing a positive pointwise mutual information between each feature and the second feature.
5. The method of claim 1 , wherein the similarity score for each feature is further generated using a term frequency-inverse document frequency between each feature and the second feature.
6. A system for generating a fusion score for electronic documents, the system comprising:
a document management system comprising at least one memory storing executable instructions and a processor configured to execute instructions, which cause the processor to perform operations comprising:
receive a first electronic document;
extract a first set of features from the first electronic document, the first set of features comprising a hierarchical structure of the first electronic document and a content of the first electronic document;
receive a second electronic document;
extract a second set of features from the second electronic document, the second set of features comprising a hierarchical structure of the second electronic document and a content of the second electronic document;
generate a similarity score for each feature by comparing each feature in the first set of features to a corresponding feature in the second set of features;
generating a weighted feature, by a machine learning model, for each feature of the first electronic document and the second electronic document, the weighted feature based on feedback from a previous similarity score; and
generating a fusion score comprising a weighted combination of the similarity scores, wherein the weighted combination is based on the weighted feature, and the fusion score represents a similarity between the first electronic document and the second electronic document.
7. The system of claim 6 , wherein the operation of extracting the first set of features comprises extracting a tree path structure from the first electronic document, wherein the tree path structure comprises a plurality of nodes.
8. The system of claim 7 , wherein the tree path structure further comprises semantic data for each node of the plurality of nodes, wherein the semantic data identifies a position of each node within the tree path structure.
9. The system of claim 6 , wherein the operation of generating the similarity score for each feature comprises computing a positive pointwise mutual information between each feature and the second feature.
10. The system of claim 6 , wherein the similarity score for each feature is further generated using a term frequency-inverse document frequency between each feature and the second feature.
11. The system of claim 6 further comprising a client device communicatively coupled to the document management system and operable to communicate the first electronic document or the second electronic document to the document management system.