Contextual Graph for Unsupervised Content Summarization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing automatic data summarization methods rely on static features, which fail to capture the context in which document terms are used, leading to inaccurate summaries due to inability to distinguish between similar terms and affect the relevance of sentence similarity values.
Innovation Solution
A computer-implemented method using a contextual graph to summarize content, where natural language processing generates semantic associations and thematic relevance, constructing a context graph with nodes representing phrases and edges representing aggregated scores for structural and semantic similarity, enabling context-based ranking and summarization.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If static features (term frequency, term length, term position) are used to calculate sentence importance, then the calculation process is simple and fast, but the indexing weight fails to reflect the context with which various document terms are used, leading to inaccurate summarization
Solution Approach 1:
The patent transforms static term-based features into dynamic contextual features by calculating contextual distances between terms using cosine similarity on TF-IDF vectors. This changes the parameter representation from simple frequency counts to context-aware similarity measurements, enabling the system to capture contextual relationships while maintaining computational efficiency through vector operations
Solution Approach 2:
The patent introduces contextual distance as an intermediary metric that mediates between static term frequency and semantic meaning. By calculating the contextual distance between terms using cosine similarity on TF-IDF vectors, the system creates a bridge that connects simple frequency-based features with contextual understanding, allowing accurate contextual representation without direct complex analysis
2Device complexity
If traditional graph-based algorithms use static term frequency to establish sentence similarity, then the algorithm complexity remains low, but the ability to distinguish between various terms with similar scores is compromised
Solution Approach 1:
The patent modifies the similarity calculation parameter from static term frequency to contextual distance based on cosine similarity. This parameter change enables the algorithm to distinguish between terms that may have similar frequencies but different contextual meanings, improving term distinction accuracy while maintaining the graph-based algorithm structure
Solution Approach 2:
The patent introduces dynamic contextual distance calculations that adapt to the specific document corpus. Instead of using fixed static features, the system dynamically computes contextual relationships between terms based on their TF-IDF vectors, allowing the algorithm to adapt to different documents and improve reliability across various contexts
Data Source
AI summary
A method, system and computer-usable medium are disclosed for using a contextual graph to summarize a corpus of content. Natural Language Processing (NLP) preprocessing operations are performed on text within an input corpus to form a grammatical analysis. In turn, the grammatical analysis is used to generate semantic associations between phrases in the input corpus. The resulting semantic associations are then used to determine the thematic relevance of the individual sentences in the input corpus to form a context-based ranking. In turn, the context-based ranking is used to construct a context graph, the vertices of which are represented by phrases, and the edges are represented by an aggregate score resulting from performing calculations associated with semantic similarity of the phrases. The resulting context graph is then used to generate a content summarization for the input corpus.