Transformer Model Matching Unstructured Text to Chat Conversations

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Conventional machine learning approaches for processing natural language in online conversations are suboptimal due to their unsupervised nature and reliance on single input queries, failing to effectively match user computers to relevant chat conversations in multi-message contexts, which is crucial for applications like merchant websites and search engines.

Innovation Solution

A computer-implemented method using a trained transformer-based machine learning model, specifically a supervised RoBERTa language model, processes unstructured text inputs to match user computers to relevant online chat conversations by generating prediction outputs with probability values, allowing for continuous re-training based on feedback and improving similarity assessment techniques.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Ease of manufacture

If unsupervised text retrieval approaches are used, then the system can process natural language text without requiring labeled training data, but the accuracy of matching users to relevant conversations deteriorates

Engineering Contradiction:
Improveease of system deploymentVSAvoidmatching accuracy
Core Design Contradiction:
Ease of manufactureVSMeasurement precision

Solution Approach 1:

The system implements feedback mechanisms where user interactions and conversation data are continuously used to retrain the transformer model, improving matching accuracy over time while maintaining initial ease of deployment

Inventive Principle:
Principle #23Feedback

Solution Approach 2:

The system transitions from static unsupervised parameters to dynamic supervised parameters by continuously updating the transformer model with labeled data from user interactions, enabling progressive improvement in matching precision

Inventive Principle:
Principle #35Parameter changes

2Device complexity

If conventional unsupervised approaches are used, then the system avoids complex training requirements, but it cannot continuously improve by updating training sets to address past errors

Engineering Contradiction:
Improvetraining complexityVSAvoidcontinuous improvement capability
Core Design Contradiction:
Device complexityVSReliability

Solution Approach 1:

The system establishes continuous improvement through ongoing retraining cycles where the transformer model processes new conversation data and user feedback indefinitely, transitioning from one-time unsupervised setup to continuous supervised learning

Inventive Principle:
Principle #20Continuity of useful action

Solution Approach 2:

User interactions and conversation outcomes provide feedback loops that enable the system to identify and correct matching errors, continuously improving reliability through iterative retraining on labeled data from past errors

Inventive Principle:
Principle #23Feedback

3Productivity

If single input query classification is used, then the system simplifies processing by focusing on one query at a time, but it fails to capture the context and topic relationships in multi-message chat conversations

Engineering Contradiction:
Improveprocessing efficiencyVSAvoidcontextual understanding accuracy
Core Design Contradiction:
ProductivityVSMeasurement precision

Solution Approach 1:

The system merges multiple individual chat messages into a unified contextual representation using the transformer model, allowing the system to process conversation context as an integrated whole rather than isolated queries, thereby improving contextual understanding while maintaining processing efficiency

Inventive Principle:
Principle #5Merging (Combining)

Data Source

PatentEP4198766A1Recommending online communication groups by matching unstructured text input to conversations
Publication Date: 2023.06.21 SUPPORTIV INC
  • EP4198766A1 patent drawingFigure 1
  • EP4198766A1 patent drawingFigure 2
  • EP4198766A1 patent drawingFigure 3

AI summary

A computer joining an online chat service, based on unstructured text input, can be matched automatically under computer control to one of multiple different online chat conversations using a trained transformer-based machine learning model, training techniques, and similarity assessment techniques. Computer analysis in this manner improves the likelihood that the unstructured text input results in assigning the computer to a relevant chat conversation. Additionally, or alternatively, a dense passage retrieval machine learning model having a first encoder for resources and a second encoder for messages can automatically match relevant resources to computers or sessions based on analysis of a series of messages of an online chat conversation. In either approach, continuous re-training is supported based on feedback from a moderator computer and/or user computers.