Dialog Token Padding Structure for Multi-Turn Intent Prediction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current task-oriented dialog systems in vehicles are complex due to integrated business logic within the machine learning models, making them less accurate and inefficient in handling multi-turn conversations.
Innovation Solution
Decouple business logic from the machine learning model by using a sequence model to predict user intents and actions, and a separate intent model to determine missing information, with a data structure that normalizes utterances using padding tokens to manage conversation flow.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If business logic is integrated within the machine learning model, then the model can handle task-oriented dialog, but the model complexity increases and accuracy decreases
Solution Approach 1:
The patent segments the dialog system into separate components: a sequence model for predicting user intents and actions, and a separate intent model for determining missing information. This segmentation removes business logic from the machine learning model, reducing complexity while maintaining task-oriented dialog handling capability.
Solution Approach 2:
The patent extracts business logic from the machine learning model and places it in separate processing modules. The sequence model focuses solely on predicting user intents and actions without embedded business logic, while the intent model handles missing information determination. This extraction reduces model complexity and improves accuracy.
2Adaptability or versatility
If business logic is integrated within the machine learning model, then the model can handle task-oriented dialog, but processing efficiency decreases
Solution Approach 1:
By segmenting the system into a sequence model for intent/action prediction and a separate intent model for missing information, the patent enables more efficient processing. Each model performs its specialized function without the overhead of integrated business logic, improving overall processing efficiency while maintaining dialog handling capability.
3Adaptability or versatility
If utterances of different lengths are processed, then the system can handle varied user input, but data structure normalization becomes complex
Solution Approach 1:
The patent applies local quality by adding padding tokens specifically to sequences that need normalization. Instead of restructuring the entire data system, padding tokens are locally applied to individual utterance sequences to achieve uniform length representation, simplifying the overall data structure while maintaining the ability to handle varied user input lengths.
Data Source
AI summary
An example operation includes one or more of receiving utterances from a user via an input device, identifying a plurality of sentences spoken by the user from the utterances, converting the utterances into a sequence of tokens and storing the sequence of tokens within a data structure, wherein the storing comprises adding padding tokens to the data structure in between the sequence of tokens to normalize a structure among the plurality of sentences within the data structure, executing a machine learning model on the data structure with the added padding tokens to determine to make a prediction, removing the padding tokens from the data structure, and executing a natural language processing (NLP) model on the sequence of tokens within the data structure with the padding tokens removed to determine a response to the user.