Memory-Based Neural Network for Interpretable Question Answering
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Automated question answering systems using neural networks face high computation costs and lack interpretability, making them unsuitable for real-time applications and hindering improvements due to their 'black box' nature.
Innovation Solution
A memory-based neural network that identifies meta-evidence nodes in an embedding space to summarize training data, allowing for the generation of support vectors and selection of appropriate answers, providing explainability and improved predictive power.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If neural networks are used for automated question answering, then answer accuracy is improved, but computation cost increases significantly
Solution Approach 1:
The patent pre-computes and stores embeddings for all training examples before deployment. During inference, the system only needs to compute embeddings for new questions and perform similarity comparisons, avoiding the need to re-process training data. This preliminary action significantly reduces real-time computation costs while maintaining answer accuracy.
Solution Approach 2:
The patent creates vector representations (embeddings) as copies of the semantic meaning of training examples. Instead of storing and processing raw training data during inference, the system uses these pre-computed vector copies to efficiently compare against new questions, reducing computation cost while preserving the essential information needed for accurate answers.
2Measurement precision
If neural networks are used for automated question answering, then answer accuracy is improved, but interpretability deteriorates
Solution Approach 1:
The patent introduces similarity scores as an intermediary between the neural network's internal representations and the final answer selection. These scores provide a transparent metric that shows how closely each training example matches the new question, making the decision process interpretable while maintaining the accuracy benefits of neural network embeddings.
Solution Approach 2:
The patent makes the previously invisible internal representations visible by transforming them into interpretable similarity scores and rankings. This is analogous to making something invisible visible - the abstract vector comparisons are translated into concrete, human-understandable metrics that reveal why certain answers are selected, without compromising the neural network's accuracy.
3Reliability
If traditional neural networks are used, then predictive power is achieved, but explainability is lost
Solution Approach 1:
The patent implements a feedback mechanism where similarity scores provide explanatory information about the model's predictions. By ranking training examples according to their similarity to the new question, the system provides feedback that explains which training instances most influenced the answer selection, maintaining predictive power while adding explainability.
Solution Approach 2:
The similarity score serves as an intermediary that bridges the gap between the neural network's black-box predictions and human understanding. It translates the abstract vector space operations into interpretable metrics that explain the reasoning behind each answer, preserving predictive power while restoring explainability.
Data Source
AI summary
The present disclosure provides a memory-based neural network for question answering. Embodiments of the disclosure identify meta-evidence nodes in an embedding space, where the meta-evidence nodes represent salient features of a training set. Each element of the training set may include a questions appended to a ground truth answer. The training set may also include questions with wrong answers that are indicated as such. In some examples, a neural Turing machine (NTM) reads a dataset and summarizes the dataset into a few meta-evidence nodes. A subsequent question may be appended to multiple candidate answers to form an input phrase, which may also be embedded in the embedding space. Then, corresponding weights may be identified for each of the meta-evidence nodes. The embedded input phrase and the weighted meta-evidence nodes may be used to identify the most appropriate answer.