Topic-Based Paper Classification for Structured Research Proposals

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Existing methods for automatic literature review generation in research proposals fail to provide structured, topic-based classification of scientific papers, often resulting in monolithic reviews lacking adequate linkage to specific topics and relying on unavailable citation intents.

Innovation Solution

A two-stage approach involving a topic-based retriever (TR) model and a topic classifier (TC) model is employed to classify scientific papers based on topic relevance, using pseudo-positive and pseudo-negative reference text spans, and enhanced regular expressions to extract in-text citations.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Ease of manufacture

If existing automatic literature review generation approaches are used, then scientific articles can be summarized independently, but the reviews become monolithic and lack structured subsections linked to specific topics

Engineering Contradiction:

Improveautomation of literature review generationVSAvoidstructure of literature review

Core Design Contradiction:

Ease of manufactureVSShape

Solution Approach 1:

The patent segments the monolithic literature review into structured subsections organized by topics. The system identifies key topics from the target paper and creates separate review sections for each topic, with relevant citations grouped under corresponding topic headings. This transformation converts a single undifferentiated review into multiple focused subsections, improving readability and topical organization.

Inventive Principle:

Principle #1Segmentation

2Extent of automation

If citation text based classification is used, then papers can be classified into predefined categories, but citation text is unavailable in early stages of research proposal writing

Engineering Contradiction:

Improveautomatic classification of papersVSAvoidavailability of citation text

Core Design Contradiction:

Extent of automationVSLoss of information

Solution Approach 1:

The system performs preliminary classification of reference papers using only metadata (titles, abstracts, keywords) before citation text becomes available. By conducting automatic classification in advance based on available information, the system prepares organized reference material that can later be enhanced with full citation text when it becomes available during proposal writing.

Inventive Principle:

Principle #10Preliminary action

Solution Approach 2:

The patent uses metadata (titles, abstracts, keywords) as an intermediary to bridge the gap between unavailable citation text and the need for paper classification. These metadata elements serve as substitutes that enable automatic classification functionality even when complete citation information is not yet available.

Inventive Principle:

Principle #24Intermediary (Mediator)

3Productivity

If existing retrieval approaches relying on abstracts and titles are used, then papers can be retrieved from corpus, but adequate information is lacking for appropriate mapping to intent or topic

Engineering Contradiction:

Improveretrieval of scientific papersVSAvoidinformation for topic mapping

Core Design Contradiction:

ProductivityVSLoss of information

Solution Approach 1:

The system merges multiple information sources including paper titles, abstracts, keywords, and full text content to create comprehensive topic mappings. By combining these diverse information sources, the system overcomes the limitations of using only abstracts and titles, achieving both efficient retrieval and accurate topical classification.

Inventive Principle:

Principle #5Merging (Combining)

4Extent of automation

If monolithic extractive or abstractive reviews are generated, then literature review can be automated, but readability is adversely affected

Engineering Contradiction:

Improveautomation of literature reviewVSAvoidreadability of literature review

Core Design Contradiction:

Extent of automationVSEase of operation

Solution Approach 1:

The patent divides the literature review into segmented topical subsections rather than generating a single monolithic review. Each subsection focuses on a specific topic with relevant citations and synthesized content, making the overall review more readable and easier to navigate while maintaining full automation through AI-generated content for each segment.

Inventive Principle:

Principle #1Segmentation

Data Source

PatentUS20250355920A1Method and system for topic-based classification of scientific papers to research proposal

Publication Date: 2025.11.20 TATA CONSULTANCY SERVICES LTD

AI summary

For categorizing literature for a specific target proposal, authors face challenges while organizing their papers in diverse ways. Embodiments of the present disclosure provide method and system for classification of scientific papers to research proposal based on topic. A constructed dataset with a positive and negative reference text spans are augmented to obtain an extended dataset. Top-k chunks from a reference paper relevant to the citation text are considered as the reference text spans. A topic-based retriever model is trained on subset of the extended dataset by the positive reference text span, and the negative reference text span. A topic classifier model is trained using research proposal title, the proposal topic, and the reference text span from the reference paper to classy if the reference paper is aligned to the topic. A reference paper to be relevant to the proposal topic is classified with corresponding topics in the research proposal.