Bot Orchestrator Training Data Harmonization for Chatbot Selection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing bot orchestrators select the wrong chatbot due to unharmonized training data sets from multiple, isolated chatbot developers, leading to misinterpretation of user inputs.
Innovation Solution
Harmonize the training data set for the bot orchestrator by filtering out duplicate, similar, or unrelated expressions using mathematical distance measures like Jaccard, Euclidean, and cosine distances, then train the machine learning-based classification model with the filtered data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If training data sets from multiple isolated chatbots are aggregated to train the metabot model, then the bot orchestrator can manage multiple chatbots, but the data sets contain duplicates and unrelated expressions causing wrong chatbot selection
Solution Approach 1:
The patent applies preliminary action by filtering and harmonizing training data sets before aggregating them for metabot model training. The system pre-processes individual chatbot training data to remove duplicates and unrelated expressions (like small talk) using similarity comparison, ensuring that only relevant and unique expressions are included in the aggregated training data. This preliminary cleaning action prevents the reliability issues that would arise from aggregating unprocessed data sets.
Solution Approach 2:
The patent applies the taking out principle by extracting and removing harmful elements from the training data sets. The system identifies and removes duplicate expressions and unrelated expressions (such as small talk) from individual chatbot training data sets before aggregation. This extraction of harmful or redundant data elements ensures that the aggregated training data maintains high quality and leads to accurate chatbot selection by the bot orchestrator.
2Adaptability or versatility
If individual chatbots are designed to work in similar scenarios, then each chatbot can handle its specific tasks, but duplicate expressions appear in training data sets
Solution Approach 1:
The patent applies preliminary action by performing similarity-based filtering on training data sets before aggregation. The system compares expressions across different chatbot training data sets using similarity metrics, identifies duplicates, and removes them in advance. This preliminary action ensures that when chatbots are designed to handle similar scenarios, their training data can be successfully aggregated without redundancy, maintaining manufacturing precision of the training data.
3Adaptability or versatility
If training data sets include unrelated expressions like small talk, then chatbots can handle diverse user inputs, but the metabot model becomes less accurate in routing
Solution Approach 1:
The patent applies the taking out principle by extracting and removing unrelated expressions from training data sets. The system identifies expressions that are unrelated to the main functionality of chatbots (such as small talk) and removes them before aggregating the training data for the metabot model. This extraction ensures that the metabot model is trained only on relevant expressions, improving input routing accuracy while maintaining the chatbots' ability to handle diverse user inputs through their individual training data.
Data Source
AI summary
Disclosed herein are system, method, and computer program product embodiments for training a bot orchestrator for chatbot selection. An embodiment operates by collecting a set of expressions from training data sets of a plurality of chatbots. For each pair of expressions, the embodiment calculates a mathematical distance between an expression for a first chatbot in the respective pair of expressions and an expression for a second chatbot in the respective pair of expressions. The embodiment filters the set of expressions based on the calculated mathematical distances, thereby generating a filtered set of expressions. The embodiment then trains a machine learning-based classification model of the bot orchestrator using the filtered set of expressions.