Siamese Architecture Infusing Knowledge Graphs for Natural Language Inference
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current natural language processing systems, particularly in sentence pair tasks like natural language inference, lack the ability to understand complex reasoning and context, leading to failures in benchmarks such as adversarialNLI and breakingNLI due to a lack of commonsense understanding, and often require large amounts of data and human intervention.
Innovation Solution
A system and method that infuse knowledge graphs and language models in a Siamese architecture, where a pre-trained language model filters context for the knowledge graph and assigns node importance, enabling joint learning and inference to capture commonsense understanding without human intervention, and handles dynamic benchmarks by combining LM and KG features effectively.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If language models are used for natural language inference tasks, then the system can process text data and derive probabilistic patterns, but the system fails to understand complex reasoning and commonsense context
Solution Approach 1:
The patent combines language models and knowledge graphs into a unified Siamese architecture where both components process input sentences simultaneously and their features are fused through attention mechanisms. This merging allows the system to leverage both the probabilistic pattern recognition of LMs and the structured reasoning capabilities of KGs, resolving the contradiction between understanding language patterns and performing complex reasoning with commonsense context.
Solution Approach 2:
The patent introduces attention mechanisms as an intermediary that mediates between language model features and knowledge graph features. The attention mechanism dynamically weights and integrates information from both sources, allowing the system to selectively combine probabilistic language patterns with structured commonsense knowledge, thereby enabling both reliable commonsense understanding and adaptable reasoning.
2Reliability
If knowledge graphs are combined with language models separately, then commonsense knowledge can be incorporated, but the learning system becomes disjointed and context information is lost
Solution Approach 1:
The patent merges knowledge graphs and language models into a single Siamese network architecture where both components share the same input processing stage and their feature representations are integrated through attention mechanisms. This unified structure eliminates the disjointed learning problem by ensuring both components operate on the same contextualized representations, preventing loss of context information while maintaining comprehensive commonsense knowledge integration.
3Productivity
If text data is used to implicitly teach language models reasoning, then the models can learn from data, but obvious facts may be missed due to dataset redundancy avoidance
Solution Approach 1:
The patent uses knowledge graphs as an intermediary that explicitly encodes commonsense facts and relationships. This intermediary compensates for the limitations of implicit learning from text data by providing structured, non-redundant commonsense knowledge that fills gaps where text datasets may have omitted obvious facts. The attention mechanism then mediates between the language model's learned patterns and the explicit commonsense knowledge from the knowledge graph.
4Measurement precision
If current benchmarks are used to evaluate language models, then model performance can be measured, but the benchmarks lead to early saturation and overestimated NLU performance
Solution Approach 1:
The patent employs dynamic benchmarks that adapt to model performance levels, preventing early saturation. The Siamese architecture with attention mechanisms enables the model to continuously improve by dynamically integrating language patterns and commonsense knowledge, allowing it to surpass static benchmarks and achieve more reliable NLU performance estimates that reflect true understanding rather than memorization.
Data Source
AI summary
The present disclosure relates to a system for infusing knowledge graphs and language models (LM) for natural language sentence pair tasks, the system include a processor operatively coupled to an inference engine, the inference engine configured to receive a sentence indicative of a premise and a sentence indicative of a hypothesis, extract LM embeddings for the corresponding sentence, generate a common-sense knowledge graph for corresponding sentence and nodes are derived from the common-sense knowledge graph, assign node importance scores for each of the derived nodes, compute node parameters for the derived nodes, apply an aggregation function to generate pooled values, concatenate the LM embeddings, graph embeddings and pooled values to generate concatenated data and classify the concatenation data to indicate a relationship between the natural language inference pair.