Multi-Turn Chatbot Response Selection via Event History Mapping
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current AI-based chatbot solutions face challenges in providing coherent responses in multi-turn conversations, especially when user inputs include both speech and physical actions, due to sparsity in training data and intent-based design limitations that lock users into one context at a time, making it difficult to handle unexpected information or context switches.
Innovation Solution
A method and computing platform that processes user inputs by organizing historical conversations into inter-related data tables, using a machine learning-based statistical model to classify new inputs, and persisting a data model of observation history and events to retrieve coherent responses through event extension, creation, or aliasing, ensuring efficient and context-aware interactions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If an intent-based design is used to infer user intent and carry out scripted content, then the bot can execute predefined tasks efficiently, but the user is locked into one context at a time, limiting the ability to respond to unexpected information or context switches
Solution Approach 1:
The patent applies dynamics by making the conversation context flexible and adaptable rather than fixed. The system dynamically determines which contexts are active based on conversation history and user input, allowing the bot to switch between multiple contexts fluidly. This resolves the contradiction by enabling both efficient task execution (when a context is clearly active) and adaptability to unexpected inputs (by activating alternative contexts when needed).
Solution Approach 2:
The patent implements multi-functionality by designing a single conversation management system that can handle multiple contexts simultaneously. Instead of separate scripted systems for each intent, one universal context management mechanism handles diverse conversation scenarios, enabling the bot to respond to various types of user inputs across different contexts without being locked into a single predefined path.
2Ease of manufacture
If training data is sparse for multi-turn conversations, then the system can be trained faster and with less data, but it becomes much less likely to have examples of every possible sequence, making coherent responses unavailable
Solution Approach 1:
The patent applies preliminary action by pre-processing and organizing training data into structured conversation contexts and histories before training the model. This preparation ensures that even with limited data, the most relevant contextual patterns are captured and organized in a way that maximizes their utility during inference, improving response coherence without requiring extensive training data.
Solution Approach 2:
The system uses feedback mechanisms where conversation history and contextual information from previous turns are fed back into the model to generate coherent responses. This allows the model to maintain consistency across multi-turn conversations by continuously incorporating contextual feedback, compensating for the sparsity of training data through effective utilization of available conversational context.
3Stability of the object's composition
If a rules-based chatbot follows pre-designed rules forming a decision tree, then the bot can provide consistent responses according to predefined logic, but it cannot understand language outside of a closed set of pre-programmed recognized inputs
Solution Approach 1:
The patent merges rules-based consistency with AI-based flexibility by combining structured context management (providing consistency) with neural language understanding (providing flexibility). The system maintains predefined conversation contexts and structures for consistency while using AI models to interpret and map diverse user inputs to these contexts, enabling both stable response patterns and adaptable language understanding.
Data Source
AI summary
A method to select a response in a multi-turn conversation between a user and a conversational bot. The conversation is composed of a set of events, wherein an event is a linear sequence of observations that are user speech or physical actions. Queries are processed against a set of conversations that are organized as a set of inter-related data tables, with events and observations stored in distinct tables. As the multi-turn conversation proceeds, a data model comprising an observation history, together with a hierarchy of events determined to represent the conversation up to at least one turn, is persisted. When a new input (speech or physical action) is received, it is classified using a statistical model to generate a result. The result is then mapped to an observation in the data model. Using the mapped observation, a look-up is performed into the data tables to retrieve a possible response.