Question Routing via Clustering for QA Pipelines

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current Question and Answer (QA) systems face challenges in ensuring that testing questions are distinct from training questions and in routing questions to the appropriate domain-specific pipelines, leading to suboptimal results due to overlap and lack of specialized training in various domains.

Innovation Solution

The proposed solution involves clustering questions based on extracted features and attributes, using a clustering engine to identify matching clusters and associate them with appropriate pipelines, allowing for separate training and testing sets and routing questions to the most relevant pipeline for accurate answers.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If questions are routed to domain-specific pipelines, then answer accuracy is improved, but system complexity increases

Engineering Contradiction:
Improveanswer accuracyVSAvoidsystem complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent divides the QA system into multiple domain-specific pipelines (e.g., medical QA pipeline, legal QA pipeline, general QA pipeline) based on question clustering. Each pipeline is specialized for handling specific domains, which improves answer accuracy by ensuring domain-expert processing while maintaining manageable complexity through modular architecture

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent introduces a clustering engine as an intermediary component that sits between the input question and the domain-specific pipelines. This mediator analyzes the input question, determines its appropriate domain, and routes it to the corresponding pipeline, thereby improving accuracy without requiring each pipeline to handle all domains

Inventive Principle:
Principle #24Intermediary (Mediator)

2Adaptability or versatility

If multiple domain-specific pipelines are created, then versatility is improved, but device complexity increases

Engineering Contradiction:
Improvesystem versatilityVSAvoidsystem complexity
Core Design Contradiction:
Adaptability or versatilityVSDevice complexity

Solution Approach 1:

The patent creates a universal routing mechanism using the clustering engine that can handle multiple domains through a single unified approach. Instead of having separate routing systems for each domain, the clustering engine universally analyzes any input question and routes it to the appropriate pipeline, thereby achieving multi-domain versatility without proportionally increasing complexity

Inventive Principle:
Principle #6Universality (Multi-functionality)

3Reliability

If questions are clustered into separate sets, then testing reliability is improved, but processing time increases

Engineering Contradiction:
Improvetesting reliabilityVSAvoidprocessing time
Core Design Contradiction:
ReliabilityVSLoss of time

Solution Approach 1:

The patent performs question clustering and pipeline routing as preliminary actions before actual question answering. By pre-classifying questions into appropriate domains and routing them to the correct pipelines in advance, the system ensures reliable testing with distinct training and testing sets while minimizing processing time during the actual answer generation phase

Inventive Principle:
Principle #10Preliminary action

Data Source

PatentUS9230009B2Routing of questions to appropriately trained question and answer system pipelines using clustering
Publication Date: 2016.01.05 INTERNATIONAL BUSINESS MACHINE CORPORATION
  • US9230009B2 patent drawing
  • US9230009B2 patent drawing
  • US9230009B2 patent drawing

AI summary

Mechanisms for selecting a pipeline of a question and answer (QA) system to process an input question are provided. An input question is received and analyzed to identify at least one feature of the input question. Clustering of the input question, with one or more previously generated clusters of questions, is performed based on the at least one feature of the input question. Based on results of the clustering, a matching cluster, of the one or more previously generated clusters, is identified with which the input question is associated. A QA system pipeline associated with the matching cluster is identified and the input question is processed using the identified QA system pipeline to generate one or more candidate answers for the input question. Each cluster in the one or more previously generated clusters has an associated QA system pipeline.