Conversational System Training Augmentation with Reduction Candidates
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conversational systems, such as chatbots, struggle to effectively analyze and respond to user inputs that utilize reductions like acronyms, abbreviations, and initialisms, as these are not adequately represented in their training data, limiting their ability to understand and engage in natural conversations.
Innovation Solution
A method and system for augmenting conversational system training by generating and incorporating proposed reductions into the training set, using properties of existing objects and external search queries, followed by filtering to create a reduced set of proposed reductions for inclusion in the training data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If the training set is expanded to include more reduction forms, then the conversational system's ability to understand user inputs improves, but the complexity of data processing and training increases
Solution Approach 1:
The system performs preliminary actions by automatically generating reduction candidates from existing training data before actual training occurs. The reduction generation module creates potential reductions from entities and intents in the training set, and the filtering module pre-filters these candidates based on validity criteria, so that when training data is augmented, only high-quality reductions are included, reducing the complexity of processing invalid or erroneous reductions.
Solution Approach 2:
The system uses self-service by leveraging its own existing training data to generate reduction candidates. The reduction generation module extracts entities and intents from the current training set and automatically creates reduction forms without requiring external manual input, thereby utilizing the system's own resources to expand its capabilities.
2Manufacturing precision
If manual creation of reduction forms is performed, then accuracy of reductions improves, but time and labor resources increase
Solution Approach 1:
The system performs self-service by automatically generating reduction candidates from its own training data using the reduction generation module. This automated process extracts entities and intents and creates reduction forms without human intervention, eliminating the time and labor required for manual creation while maintaining acceptable accuracy through subsequent filtering.
Solution Approach 2:
The filtering module provides feedback by evaluating generated reduction candidates against predefined criteria and filtering out invalid or low-quality reductions. This feedback mechanism ensures that only accurate reductions are added to the training set, maintaining high precision without requiring manual verification of each candidate.
3Adaptability or versatility
If all generated reduction candidates are included in training, then coverage of reduction forms improves, but data quality and relevance decrease
Solution Approach 1:
The filtering module implements feedback by systematically evaluating each reduction candidate against validity criteria such as whether the reduction corresponds to an existing entity or intent, whether it follows proper reduction patterns, and whether it is relevant to the domain. This feedback loop ensures that only high-quality, relevant reductions are included in the augmented training set, maintaining data reliability while still achieving broad coverage.
Solution Approach 2:
The system changes parameters by adjusting the filtering criteria and thresholds to balance coverage and quality. By modifying the strictness of validation rules and the criteria for inclusion, the system can dynamically control the trade-off between how many reduction forms are included and how high the overall quality standard remains.
Data Source
AI summary
A method, system and computer-usable medium for augmenting the training of a conversational system. In certain embodiments, the method comprises: ingesting a training set to be used in training of the conversational system, the training set including objects for use in the training, wherein the objects include one or more object types, wherein the object types include one or more of an entity or intent; generating proposed reductions for inclusion in an augmented training set, wherein the proposed reductions include one or more of: reduction candidates generated using properties of objects included in the training set; reduction candidates obtained from search queries of one or more external resources, wherein the search queries relate to one or more objects included in the training set; filtering the proposed reductions to generate a reduced set of proposed reductions; and augmenting the training set with the reduced set of proposed reductions.