Multi-Round Machine Reading Comprehension for Causal Extraction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for extracting causal relationships from text, such as rule-based and deep learning models, face challenges in accuracy due to limitations in handling complex causality, multiple pairs of causes and effects, and implicit causality, leading to low extraction accuracy.
Innovation Solution
A multi-round multi-answer machine reading comprehension model is employed, utilizing two machine reading comprehension models to iteratively extract causes and effects from text, improving accuracy through sequential processing and incorporating knowledge-enhanced vectors for better handling of complex causality.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Extent of automation
If rule-based or deep learning models are used for causal relationship extraction, then the extraction process can be automated, but the accuracy is low due to limitations in handling complex causality, multiple pairs of causes and effects, and implicit causality
Solution Approach 1:
The patent segments the causal relationship extraction task into multiple rounds, where each round focuses on extracting specific cause-effect pairs. The multi-round extraction process divides complex causality into manageable iterations, with each round targeting particular causal relationships identified through question generation and answer extraction mechanisms.
Solution Approach 2:
The patent introduces an intermediary question generation mechanism that creates specific questions about causes and effects as a bridge between the text and the extraction model. These generated questions act as mediators that guide the extraction process, enabling the model to focus on specific causal relationships rather than attempting to extract all relationships in a single pass.
2Productivity
If a single-round extraction method is used, then the processing speed is fast, but the accuracy is low due to inability to handle complex causality and multiple cause-effect pairs
Solution Approach 1:
The patent implements periodic action through multi-round extraction, where the extraction process is repeated in multiple iterations rather than executed once. Each round extracts specific cause-effect pairs, and the process periodically refines results by generating new questions and extracting additional relationships in subsequent rounds, balancing speed and accuracy through structured repetition.
Solution Approach 2:
The patent applies preliminary action by generating questions about causes and effects before the actual extraction process. These pre-generated questions serve as preparation that guides the extraction model, enabling more accurate and targeted extraction in the main processing phase rather than attempting to extract all relationships simultaneously without prior guidance.
3Device complexity
If traditional machine reading comprehension models are used, then the model structure is simple, but the model cannot effectively handle implicit causality and complex causal relationships
Solution Approach 1:
The patent introduces dynamics by making the question generation process adaptive based on the text content. The model dynamically generates questions about causes and effects specific to each text, allowing the extraction process to adapt to different types of causal relationships including implicit causality, rather than relying on fixed static question templates.
Solution Approach 2:
The patent adds another dimension to the extraction process by introducing a question generation layer that operates in a different dimensional space. Instead of directly extracting cause-effect pairs from text, the model first generates questions in a question space, then uses these questions to guide extraction, effectively adding a dimensional transformation that enhances the model's ability to handle complex causality.
Data Source
AI summary
A method is provided that includes: obtaining a text; inputting the text and a first question into a first machine reading comprehension model to obtain a first round of answer, wherein the first question inquires about cause of a causal relationship in the text, and the first round of answer includes at least one cause; and inputting the text and a second question into a second machine reading comprehension model to obtain a second round of answer, wherein the second question inquires about effect of the causal relationship in the text, and the second round of answer includes at least one effect of the at least one cause.