TRIZ-RAGNER: A Retrieval-Augmented Large Language Model for TRIZ-Aware Named Entity Recognition in Patent-Based Contradiction Mining

1. Problem Statement

The paper addresses the challenge of TRIZ-based contradiction mining in patent analysis. The goal is to automatically identify pairs of technical parameters from patent texts where improving one parameter leads to the deterioration of another (a "technical contradiction").

• Limitations of Existing Methods:
  • Rule-based systems: Lack scalability and robustness against diverse patent expressions.
  • Traditional ML/Deep Learning (e.g., BiLSTM-CRF, BERT): Rely on surface-level token labeling and struggle with semantic ambiguity, long-distance dependencies, and implicit technical semantics. They often fail to generalize across different patent domains.
  • Direct LLM Application: While Large Language Models (LLMs) possess strong semantic understanding, they suffer from hallucinations and a lack of grounding in structured TRIZ knowledge, leading to inconsistent or incorrect parameter extraction.

2. Methodology: TRIZ-RAGNER Framework

The authors propose TRIZ-RAGNER, a framework that reformulates contradiction mining as a semantic-level Named Entity Recognition (NER) task. Instead of labeling tokens, the model extracts conceptual entities (Improving and Worsening Parameters) by grounding LLM reasoning in external TRIZ knowledge. The framework operates in three main stages:

A. TRIZ Knowledge Base Construction

A structured knowledge base ($K$) is built containing:

• The 39 standard TRIZ engineering parameters and their definitions.
• Annotated examples from the PaTRIZ dataset.
• Extended domain knowledge linking parameters to real engineering scenarios.
• Each entry is encoded into dense vectors for semantic search.

B. Retrieval and Reranking Pipeline

1. Dense Retrieval: The input patent sentence is encoded into a vector. A dense retriever (using cosine similarity) fetches the top-$K$ most relevant TRIZ knowledge snippets from the knowledge base.
2. Cross-Encoder Reranking: To mitigate noise from coarse retrieval, a cross-encoder model ($g$) jointly encodes the input sentence and each candidate knowledge snippet. It computes a relevance score to filter and refine the context, ensuring only highly pertinent TRIZ semantics are passed to the LLM.

C. Structured Prompting and LLM Inference

• A structured prompt is constructed by concatenating the input sentence, the refined knowledge context, and a specific instruction template.
• The LLM is instructed to output the result in a strict JSON-like format (e.g., {"Improving_Parameter": ..., "Worsening_Parameter": ...}).
• This structured prompting constrains the generation space, reducing ambiguity and ensuring the output adheres to the required entity schema.

3. Key Contributions

1. Semantic-Level Reformulation: The authors shift the paradigm from traditional token-level sequence labeling to a semantic-level NER task, treating TRIZ parameters as conceptual entities rather than surface spans.
2. Retrieval-Augmented Framework: They introduce TRIZ-RAGNER, which integrates Retrieval-Augmented Generation (RAG) to inject domain-specific TRIZ knowledge directly into the LLM inference process, effectively reducing hallucinations.
3. Cross-Encoder Reranking: A novel mechanism is introduced to filter retrieved knowledge, ensuring high relevance and reducing semantic noise before the LLM processes the context.
4. Comprehensive Evaluation: Extensive experiments on the PaTRIZ dataset demonstrate the framework's superiority over both traditional sequence labeling models and prompt-engineered LLM baselines.

4. Experimental Results

The model was evaluated on the PaTRIZ dataset (approx. 3,000 annotated patent sentences) using Precision, Recall, and F1-score.

• Performance Metrics:
  • Precision: 85.6%
  • Recall: 82.9%
  • F1-Score: 84.2%
• Comparisons:
  • Outperformed traditional models (BiLSTM-CRF, BERT-CRF, RoBERTa-CRF) which achieved F1-scores between 60.7% and 71.7%.
  • Outperformed the strongest baseline (GPT with prompt engineering), which achieved an F1-score of 76.9%.
  • Improvement: TRIZ-RAGNER achieved an absolute 7.3 percentage point improvement in F1-score over the best baseline.
• Ablation Study:
  • Removing the Retrieval Module dropped F1 to 78.1%.
  • Removing Reranking dropped F1 to 79.8%.
  • Removing Structured Prompting dropped F1 to 76.6%.
  • Conclusion: All components (retrieval, reranking, and prompting) are essential and complementary.

5. Significance and Impact

• Robustness in Technical Domains: The study demonstrates that grounding LLMs with structured, domain-specific knowledge (TRIZ) is critical for accurate information extraction in complex engineering texts where "hallucination" is a major risk.
• Systematic Innovation: By automating the extraction of contradiction pairs, the framework accelerates the process of computer-aided inventive design, helping engineers uncover innovation opportunities and apply TRIZ principles more effectively.
• Future Directions: The authors note that while the current approach is effective, it introduces computational overhead. Future work will focus on adaptive retrieval strategies and lightweight reranking mechanisms to improve efficiency, as well as expanding to document-level and multilingual mining.

Note on Figure 2: The provided text contains a description of Figure 2 regarding "cumulative systemic loss ratio" and "SR-DTMA," which appears to be a copy-paste error from a different paper (likely related to Digital Twins and Reinforcement Learning). This content is irrelevant to the TRIZ-RAGNER paper and has been excluded from this summary.