Fusion Model for Natural Language Text Generation from Structured Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current data-to-text generation methods, such as template-based and sequence-to-sequence with attention approaches, face challenges in producing fluent and grammatical sentences while being computationally extensive and lacking user control.
Innovation Solution
A fusion model integrating a sentence generation component and a template generation component, using Bidirectional Long Short-Term Memory (Bi-LSTM) encoders and decoders, generates natural language text by processing input dictionaries through attention mechanisms to produce candidate sentences and templates, with a ranking system to select the most probable sentences based on fidelity to the input data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of operation
If a template-based approach is used for data-to-text generation, then the generation process is simple and controllable, but the generated sentences may not be fluent or grammatical and the approach requires very large and diverse training data sets which is laborious
Solution Approach 1:
The patent combines the template-based approach and sequence-to-sequence approach into a fusion model. The model integrates a template generation component that provides controllability with a sentence generation component using Bi-LSTM and attention mechanisms that ensure fluency and grammaticality, resolving the contradiction between ease of operation and reliability
2Reliability
If a sequence-to-sequence with attention approach is used for data-to-text generation, then fluent and grammatical sentences can be generated, but the approach is computationally extensive and lacks user control
Solution Approach 1:
The patent segments the sequence-to-sequence model into two distinct components: a template generation component that provides structural guidance and a sentence generation component that produces fluent text. This segmentation allows the model to use computational resources more efficiently by separating the control function from the generation function
Solution Approach 2:
The patent introduces sentence templates as an intermediary between the input structured data and the output natural language text. These templates serve as a bridge that guides the generation process, reducing the computational burden while maintaining fluency and grammaticality
3Reliability
If a sequence-to-sequence with attention approach is used for data-to-text generation, then fluent sentences can be generated, but the approach lacks user control
Solution Approach 1:
The patent merges the controllability of template-based methods with the fluency of sequence-to-sequence methods in a unified fusion model. The model allows users to control the generation process through template selection while maintaining natural language fluency through the integrated neural network components
Data Source
AI summary
Generating natural language text from structured data using a fusion model is disclosed. Based on an input dictionary, a first sequence of vectors is generated by a first encoder and a second sequence of vectors is generated by a second encoder. The first and second sequences of vectors are provided to an attention function which generates a modified sequence of vectors. A decoder decodes the modified sequence of vectors to generate a plurality of ordered sequences corresponding to a target natural language sentence. A predetermined number of candidate sentences are determined based on the plurality of ordered sequences and are ranked to select a sentence as the target natural language sentence.