Automatic Paraphrase Generation for NLU Training Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Creating training data for conversational agents is time-consuming, especially for new products or features, as it requires generating and labeling utterances to understand user commands, which is a bottleneck in extending Natural Language Understanding (NLU) models.
Innovation Solution
An automatic paraphrase generation system based on machine translation is used to generate semantically equivalent utterances, allowing developers to quickly create diverse training data by selecting from automatically generated paraphrases, reducing the effort required for NLU model creation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If manual utterance generation and labeling is performed to create training data for NLU models, then the quality and accuracy of training data can be ensured, but the time consumption and labor effort increase significantly
Solution Approach 1:
The system generates multiple paraphrase versions of each training utterance automatically, creating copies with varied expressions but identical semantic meaning. This allows the training data to be expanded without manual creation of each example, reducing time consumption while maintaining quality through the use of established paraphrase generation models
Solution Approach 2:
The system enables developers to self-generate diverse training data by simply selecting from automatically generated paraphrases rather than manually creating each utterance. The paraphrase generation process is automated and self-service oriented, allowing rapid expansion of training datasets without proportional increase in manual effort
2Adaptability or versatility
If diverse training utterances are manually created to improve NLU model coverage, then the model's understanding of various user intents improves, but the complexity and effort of data creation increases
Solution Approach 1:
The paraphrase generation system serves multiple functions: it generates diverse utterances for training data, provides semantic variations for better model coverage, and maintains consistency through controlled generation. This single system addresses multiple needs in training data creation, reducing overall complexity
Solution Approach 2:
The system varies parameters such as sentence structure, vocabulary choices, and expression patterns while maintaining the core semantic meaning. This allows generation of diverse utterances that improve model coverage without requiring complex manual intervention for each variation
3Reliability
If real user interaction data is collected to train NLU models, then the training data reflects actual usage patterns, but this approach is not available for new products or features without existing users
Solution Approach 1:
The system generates training data in advance using automated paraphrase generation, allowing new products and features to have training data available before actual user interactions occur. This preliminary action enables parallel development of NLU models alongside product development, improving productivity without sacrificing the quality of training data
Data Source
AI summary
The present teaching relates to generating a conversational agent. In one example, a plurality of input utterances may be received from a developer. A paraphrase model is obtained. The paraphrase model is generated based on machine translation. For each of the plurality of input utterances, one or more paraphrases of the input utterance are generated based on the paraphrase model. For each of the plurality of input utterances, at least one of the one or more paraphrases is selected based on an instruction from the developer to generate selected paraphrases. The conversational agent is generated based on the plurality of input utterances and the selected paraphrases.