Neural Pipeline for Entity Linking Accuracy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current systems face challenges in accurately answering simple questions in natural language due to the variety of ways questions can be phrased, despite the availability of rich knowledge graphs, as they often rely on holistic chunk generation and entity linking rather than identifying the entity of interest and utilizing relations to disambiguate entities.
Innovation Solution
A neural pipeline system that uses sequential labeling with Stacked Bidirectional Gated-Recurrent-Units (GRU) to identify the subject string, followed by relation ranking and joint disambiguation, which reduces the number of candidate entities and leverages the context of the question to determine the correct subject and relation, utilizing continuous embeddings and knowledge graphs to provide accurate answers.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If holistic chunk generation and entity linking are used, then the system can process questions, but the accuracy deteriorates due to noise and inability to identify the entity of interest
Solution Approach 1:
The patent segments the question processing into distinct components: subject identification, relation identification, and entity linking. By using sequential labeling to identify the subject chunk and relation type separately before performing entity linking, the system avoids the noise of holistic chunk generation and focuses on the specific entity of interest, thereby improving entity linking accuracy while reducing information loss.
Solution Approach 2:
The patent introduces an intermediary subject identification step that acts as a mediator between the raw question and the entity linking process. By first identifying the subject chunk and relation type as intermediate representations, the system filters out irrelevant information before entity linking, reducing noise in candidate entities while maintaining processing capability.
2Adaptability or versatility
If the system considers all candidate entities, then it maintains completeness, but the complexity increases due to the tremendous variety of question phrasings
Solution Approach 1:
The patent divides the complex task of handling question variety into segmented subtasks: subject identification using sequential labeling, relation type classification, and entity linking. This segmentation allows the system to adapt to different question phrasings by focusing on identifying key components (subject and relation) rather than processing all possible candidate entities, thereby maintaining versatility while reducing processing complexity.
Solution Approach 2:
The patent changes the parameter of processing from considering all candidate entities to focusing on identified subject-relation pairs. By using sequential labeling to identify the subject chunk and relation type as key parameters, the system adapts to question variety through these identified parameters rather than exhaustively processing all possible entities, reducing complexity while maintaining adaptability.
3Measurement precision
If sequential labeling with Stacked Bidirectional GRU is used, then subject identification accuracy improves, but the processing time increases
Solution Approach 1:
The patent applies preliminary action by using sequential labeling with Stacked Bidirectional GRU to identify the subject chunk and relation type before performing entity linking. This preliminary identification of key components (subject and relation) achieves high accuracy in subject identification, and by focusing subsequent processing only on relevant candidate entities matching the identified subject-relation pair, the system mitigates the time cost of the complex preliminary processing.
Data Source
AI summary
Described herein are systems and methods for determining how to automatically answer questions like “Where did Harry Potter go to school?” Carefully built knowledge graphs provide rich sources of facts. However, it still remains a challenge to answer factual questions in natural language due to the tremendous variety of ways a question can be raised. Presented herein are embodiments of systems and methods for human inspired simple question answering (HISQA), a deep-neural-network-based methodology for automatic question answering using a knowledge graph. Inspired by human's natural actions in this task, embodiments first find the correct entity via entity linking, and then seek a proper relation to answer the question—both achieved by deep gated recurrent networks and neural embedding mechanism.