Story Detection Using Event and Character Features
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing story detection methods in natural language processing face challenges in efficiently separating stories from non-story text, relying on manual curation, complex preprocessing pipelines, and lacking features that capture the essence of stories, leading to poor generalizability and performance across different corpora.
Innovation Solution
A story detection system utilizing 283 features that focus on events and characters, including verb-based features from VerbNet and co-reference chains, processed through a feature extraction pipeline and supported by a support vector machine for accurate classification, achieving improved performance and generalizability.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of information
If tens of thousands of features and complicated pre-processing pipelines are used for story detection, then feature comprehensiveness is improved, but device complexity increases
Solution Approach 1:
The patent extracts only the essential features needed for story detection (283 features focusing on events and characters) from the vast space of possible features, eliminating unnecessary complexity while retaining discriminative power. This selective extraction approach maintains feature comprehensiveness for story detection without requiring tens of thousands of features.
Solution Approach 2:
The patent develops a universal story detection system that generalizes across different corpora (blog posts, news articles, social media) using a consistent feature set and classification approach. The 283 features are designed to capture universal story elements (events, characters, narratives) that appear across diverse text types, eliminating the need for corpus-specific preprocessing pipelines.
2Measurement precision
If manual curation of story data sets is used, then data quality is improved, but productivity decreases
Solution Approach 1:
The patent employs automated feature extraction and classification systems that process large volumes of text data without manual intervention. The system automatically identifies story elements, extracts features, and classifies text as story or non-story, enabling high-productivity processing of electronic stories while maintaining consistent quality through algorithmic rigor.
Solution Approach 2:
The patent transforms the data curation process from manual annotation to automated feature-based classification. By changing the operational parameters from human expert review to machine learning classification using 283 structured features, the system achieves both high throughput (processing daily generated stories) and maintained quality through consistent application of detection rules.
3Reliability
If domain-specific training data is used, then training effectiveness is improved, but adaptability decreases
Solution Approach 1:
The patent creates a universal story detection model trained on diverse corpora (blog posts, news articles, social media) that can generalize across different domains. The 283 features are designed to capture universal narrative elements that appear across all text types, enabling the system to maintain high effectiveness when trained on one corpus and perform well when tested on different corpora without retraining.
Solution Approach 2:
The patent performs preliminary feature extraction and analysis across multiple corpora to identify universal story patterns before final model training. By pre-identifying the 283 critical features that appear across diverse text types, the system prepares a robust feature set that enables effective training on any corpus while maintaining adaptability to new domains.
4Device complexity
If 283 features are used instead of tens of thousands, then device complexity is reduced, but measurement precision may worsen
Solution Approach 1:
The patent extracts the 283 most discriminative features for story detection from the vast space of possible features, focusing on event-based and character-based features that are essential for identifying narrative structure. This selective extraction maintains measurement precision by retaining only the features that directly contribute to story detection accuracy while eliminating redundant features.
Solution Approach 2:
The patent applies different feature types with appropriate weighting based on their local importance for story detection. Event-based features (verbs, actions) and character-based features (entities, relationships) are extracted and weighted according to their specific contribution to narrative identification, ensuring that each feature contributes optimally to overall detection accuracy without requiring all possible features.
Data Source
AI summary
Methods and devices for story detection in text are provided. A device can include an input device for receiving text data a processor configured to: tokenize each paragraph in the text data and split each tokenized paragraph into sentences; parse each sentence from the tokenized paragraphs; label each predicate in each sentence with its respective semantic role; and assign a verb class to each predicate; and determine whether respective arguments of each predicate contains a character. The device can further include a support vector machine configured to determine whether a story is present within each paragraph based upon whether each predicate contains a character.