Dialogue Cache Generation Using Keyword-Indexed Utterance Pairs
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing dialogue systems face high computational loads and inefficiencies in creating and updating cache data due to the need to process large-scale passage sets and reflect daily-updated information, leading to excessive processing demands.
Innovation Solution
A method involving an utterance data generating device that uses trained word sequence generation models to create cache data efficiently by generating word sequence pairs, including a key word sequence for retrieval, and storing these pairs in a manner that allows for efficient reading and updating.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If a large-scale passage set is used to generate responses, then the system can provide proper responses across various domains, but the processing load increases significantly
Solution Approach 1:
The patent pre-generates question-answer pairs from the large-scale passage set and stores them in cache data before actual dialogue processing. When a user utterance arrives, the system first checks if a pre-generated question matches the utterance, avoiding the need to process the entire passage set in real-time. This preliminary preparation resolves the contradiction by shifting computational work from online to offline.
Solution Approach 2:
The patent extracts and stores only the essential question-answer pairs from the large-scale passage set into a compact cache structure. Instead of searching through all passages for each query, the system extracts relevant Q&A pairs in advance and stores them with associated topic word sequences as indexes, reducing the processing burden while maintaining comprehensive domain coverage.
2Loss of time
If cache data is created to reduce processing load, then response time decreases, but the computational cost to create and update cache data increases
Solution Approach 1:
The patent generates cache data for only a subset of topic word sequences rather than attempting to pre-process all possible queries. By selectively creating cache entries for frequently occurring or important topics, the system reduces the computational cost of cache creation while still achieving significant response time improvements for common queries.
Solution Approach 2:
The system automatically updates cache data when new passages are added to the passage set, without requiring manual intervention or complete re-processing. The update mechanism selectively generates cache entries for new topics and integrates them into the existing cache structure, reducing the computational overhead of cache maintenance while keeping the cache current.
3Reliability
If deep learning-based processes are executed in parallel to handle complex tasks, then response quality improves, but computational resources and processing time increase
Solution Approach 1:
The patent divides the complex dialogue processing task into separate stages: cache matching using simple topic word sequence comparison, and full deep learning-based question answering only when cache matching fails. This segmentation allows the system to use minimal computational resources for most queries while maintaining the option to invoke resource-intensive processes only when necessary for complex queries.
Data Source
AI summary
An utterance data generating device providing a dialogue device, a training device and an utterance data generating device that enable highly efficient generation of cache data in a dialogue device, includes: a cache data generating device generating, from each of a plurality of passages, cache data including an utterance word sequence forming a response utterance to an input utterance and a key word sequence to be a key for searching for an utterance word sequence; and a cache data storage device storing the cache data generated by the cache data generating device in a manner at least allowing reading by using the key word sequence as a key.