Interpretable Neural Text Generation Through Substring Alignment
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional neural networks in natural language processing lack interpretability, making it difficult for users to understand the process by which output text is generated from input text, particularly in terms of the correspondence between substrings in the input and output text.
Innovation Solution
A neural network apparatus that outputs both predicted text and alignment information, indicating the basis of information for each part of the input text in the generation of the output text, allowing for higher interpretability and easier correction of the output.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of operation
If conventional neural networks are used for natural language processing, then text generation capability is achieved, but interpretability of the generation process deteriorates
Solution Approach 1:
The patent introduces alignment information as an intermediary element that mediates between the complex neural network processing and the user's understanding. This alignment information maps input substrings to output substrings, providing an interpretable bridge that reveals the generation process without simplifying the underlying neural network architecture.
Solution Approach 2:
The patent segments both the input and output texts into substrings and establishes correspondence relationships between them. By dividing the continuous text into discrete segments and tracking their transformations, the system makes the internal processing of the neural network visible and understandable through alignment information.
2Loss of information
If alignment information is added to neural network output, then interpretability is improved, but information processing complexity increases
Solution Approach 1:
The alignment information serves as feedback that returns to the user about the neural network's internal decision-making process. This feedback loop provides transparency by showing which input substrings correspond to which output substrings, allowing users to verify and understand the generation process without adding complex processing to the neural network itself.
Data Source
AI summary
Disclosed is a natural language processing technique according to a neural network of high interpretive ability. One embodiment of the present disclosure relates to an apparatus including a trained neural network into which first natural language text is input and that is trained to output second natural language text and alignment information, the second natural language text being in accordance with a predetermined purpose corresponding to the first natural language text, and the alignment information indicating, for each part of the second natural language text, which part of the first natural language text is a basis of information for generation; and an analyzing unit configured to output, upon input text being input into the trained neural network, a predicted result of output text in accordance with a predetermined purpose, and alignment information indicating, for each part of the predicted result of the output text, which part of the input text is a basis of information for generation.