Implicit Rhetorical Relation Prediction Using Simplified Feature Sets
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing systems for predicting implicit rhetorical relations in large-scale annotated corpora face significant challenges, achieving only mediocre performance with current feature sets and methods, particularly in the absence of explicit discourse markers, where the task becomes computationally difficult and achieves macro-F1 scores of around 36-40% at best.
Innovation Solution
A simplified feature set based on raw text and semantic dependencies is used, including text unigrams, bigrams, trigrams, and dependency parses, to predict implicit rhetorical relations, which meets or exceeds previous performance by up to 5% for all relations and up to 14% for certain individual relations, indicating that more data may not necessarily improve performance with these features.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If complex feature sets are used to predict implicit rhetorical relations, then prediction accuracy may improve, but computational difficulty and processing complexity increase significantly
Solution Approach 1:
The patent extracts and selects only the most essential features from the text (unigrams, bigrams, trigrams, and dependency parses) while discarding redundant or less informative features. This extraction approach achieves state-of-the-art prediction accuracy using a streamlined feature set, thereby reducing computational complexity while maintaining high measurement precision in implicit rhetorical relation prediction.
2Measurement precision
If more data is collected and used for training, then prediction performance may improve, but processing time and computational resources increase
Solution Approach 1:
The patent changes the parameter of feature representation by using simplified textual features (unigrams, bigrams, trigrams, and dependency parses) instead of more complex feature representations. This parameter change enables the system to achieve high prediction performance on the Penn Discourse Treebank corpus while reducing the computational resources and processing time required for training and inference.
3Ease of operation
If explicit discourse markers are used to indicate rhetorical relations, then prediction becomes easier, but the task becomes less challenging and less applicable to real-world text without markers
Solution Approach 1:
The patent enables the system to self-determine rhetorical relations by analyzing the textual content and structural dependencies itself, without relying on explicit discourse markers. The use of unigrams, bigrams, trigrams, and dependency parses allows the system to infer rhetorical relations autonomously from the text structure and word co-occurrence patterns, making it adaptable to both marked and unmarked text while maintaining high prediction accuracy.
Data Source
AI summary
The present invention provides a method and system directed to predicting implicit rhetorical relations between two spans of text, e.g., in a large annotated corpus, such as the Penn Discourse Treebank (“PDTB”), Rhetorical Structure Theory corpus, and the Discourse Graph Bank, and particularly directed to determining a rhetorical relation in the absence of an explicit discourse marker. Surface level features may be used to capture pragmatic information encoded in the absent marker. In one manner a simplified feature set based only on raw text and semantic dependencies is used to improve performance for all relations. By using surface level features to predict implicit rhetorical relations for the large annotated corpus the invention approaches a theoretical maximum performance, suggesting that more data will not necessarily improve performance based on these and similarly situated features.