Abstractive Summary Generation via Semantic Graph Indexing
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing abstractive summary generation techniques fail to effectively retrieve and process sentences based on semantic graphs and graph index data, leading to unclear summaries due to the inability to identify and remove redundant information.
Innovation Solution
A method and system that receive a query, split it into lexical units, generate a semantic graph and graph index, retrieve sentences with analogous semantic graphs, and re-order or re-phrase them based on information measures to generate clear and concise abstractive summaries.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of information
If existing abstractive summary generation techniques are used, then summary generation is achieved, but the summary lacks clarity due to inability to retrieve and remove redundant sentences
Solution Approach 1:
The patent segments the summary generation process into distinct modules: query processing module that splits queries into lexical units, semantic graph generation module that creates structured representations, sentence retrieval module that fetches relevant sentences, and redundancy removal module that eliminates duplicate information. This segmentation allows each module to specialize in one aspect, improving overall clarity while managing complexity through modular design.
Solution Approach 2:
The patent introduces semantic graphs as an intermediary data structure between the query and the summary generation process. These graphs serve as a mediator that structurally represents the semantic relationships in the query, enabling precise sentence retrieval and redundancy identification. The graph index acts as another intermediary that facilitates efficient matching between query semantics and document sentences.
2Measurement precision
If semantic graph-based sentence retrieval is implemented, then relevant sentences are accurately identified, but the processing time increases
Solution Approach 1:
The patent performs preliminary actions by pre-processing the document into semantic graphs and building a graph index before the actual summary generation query is executed. This pre-computation of semantic structures and indexing allows the system to quickly match incoming queries against pre-built representations, achieving high retrieval precision without incurring the full computational cost during query time.
Solution Approach 2:
The patent replaces traditional mechanical text-matching systems with semantic graph-based representation and matching. Instead of relying on keyword overlap or simple string matching, the system uses graph structure comparison and semantic role labeling to identify relevant sentences, achieving higher precision through semantic understanding rather than mechanical pattern matching.
3Loss of information
If multiple sentences conveying same information are not removed, then all information is retained, but the summary becomes redundant and unclear
Solution Approach 1:
The patent implements feedback mechanisms where the semantic graph representation of the query is compared against semantic graphs of retrieved sentences. This comparison provides feedback on which sentences truly match the query intent and which are redundant. The system uses this feedback to selectively retain or remove sentences, ensuring clarity while maintaining efficiency in information delivery.
Solution Approach 2:
The patent changes the parameter of sentence selection from simple keyword matching to semantic graph similarity measurement. By transforming the selection criterion into a different parameter space (semantic structure rather than textual overlap), the system can more effectively identify and remove redundant sentences while preserving those that provide unique information, thereby improving summary clarity.
Data Source
AI summary
The present invention relates to a method for generating abstractive summary. The method comprises receiving a query for generating an abstractive summary from a document and splitting the query into one or more lexical units. Further, a semantic graph and a graph index is generated based on a role assigned to the one or more lexical units. Furthermore, a measure of information is determined for the retrieved one or more sentences. The one or more sentences having a semantic graph analogous to the generated semantic graph of the query are retrieved from the document. Finally, at least one of re-ordering and re-phrasing is performed on at least one of the retrieved one or more sentences based on the computed measure of information and the one or more lexical units in the retrieved one or more sentences to generate the abstractive summary.