Conversational AI Response Generation Using Extractive and Generative Neural Networks
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conversational artificial intelligence systems often provide non-conversational responses, which can lead to an unpleasant user experience, and existing methods for fine-tuning generative models are resource-intensive and time-consuming, limiting their applicability.
Innovation Solution
A pipeline combining an extractive question answering (EQA) model with a zero-shot generative answer extender (AE) model to generate conversational responses, where the EQA model retrieves answers and the AE model reformulates them to create natural-sounding interactions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If fine-tuning generative models is used to improve conversational responses, then response quality is improved, but resource consumption and training time increase
Solution Approach 1:
The system divides the response generation task into two separate models: an extractive question answering model for retrieving accurate information and a generative answer extender model for forming conversational responses. This segmentation allows each model to be specialized and optimized independently, reducing the overall resource requirements compared to fine-tuning a single large generative model.
Solution Approach 2:
The extractive question answering model acts as an intermediary between the user query and the generative answer extender model. It retrieves precise answers from source text, which are then passed to the generative model for conversational formatting. This intermediary approach ensures accuracy while reducing the burden on the generative model.
2Reliability
If fine-tuning generative models is used to improve conversational responses, then response quality is improved, but training time increases
Solution Approach 1:
By segmenting the task into extraction and generation components, the system avoids the lengthy fine-tuning process required for single-model approaches. The extractive model can be trained independently on extraction tasks, and the generative model on conversational formulation, significantly reducing total training time.
Solution Approach 2:
The extractive question answering model performs preliminary action by retrieving accurate answers before the generative model formats them. This preliminary extraction step ensures that the generative model only needs to learn conversational formatting rather than both information retrieval and response generation, reducing training requirements.
3Ease of operation
If generative models are used to provide conversational responses, then natural-sounding responses are generated, but information accuracy decreases due to hallucination
Solution Approach 1:
The extractive question answering model serves as an intermediary that guarantees information accuracy by retrieving answers directly from source text. This intermediary step prevents the generative model from hallucinating information, as it only generates responses based on extracted facts.
Solution Approach 2:
The system uses itself to ensure accuracy: the extractive model extracts information that is then used by the generative model. This self-service mechanism where the system's own extraction capability guarantees the accuracy of generated responses eliminates the need for external verification.
4Reliability
If extractive question answering models are used, then information accuracy is improved, but response naturalness decreases
Solution Approach 1:
The system merges the extractive question answering model with the generative answer extender model in a pipeline architecture. The extractive model provides accurate information, and the generative model adds conversational naturalness, combining the strengths of both approaches into a unified system.
Solution Approach 2:
The generative answer extender model acts as an intermediary that transforms the accurate but non-conversational extracted answers into natural-sounding responses. It mediates between the extractive model's precise output and the user's expectation for conversational interaction.
Data Source
AI summary
Systems and methods determine an answer to an input query and provide a conversational response. The answer may be determined using a trained first neural network to extract the answer from a corpus of information. The answer and the input query may be provided to a second trained neural network to generate a formulation of the input query combined with the answer in order to generate a conversational response.