Key-Phrase Extraction for Noisy Live Transcript Domains
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine learning systems for key-phrase extraction are not well-suited for noisy live transcripts, often extract redundant phrases, and lack domain-specific training for real-time transcription contexts.
Innovation Solution
A deep-learning based key-phrase extraction network trained with a domain discriminator network to identify domain-relevant key-phrases and generate meta-data, capable of handling noisy transcripts and ensuring uniqueness and domain-specific relevance.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If existing machine learning systems are trained on edited and grammatically correct documents, then they achieve good extraction performance on such documents, but they fail to efficiently and accurately extract key-phrases from noisy live transcriptions
Solution Approach 1:
The patent changes the training data parameters from edited documents to noisy live transcriptions, and introduces domain discriminator networks that dynamically adjust model parameters based on the detected domain (scientific article vs. live transcript). This allows the system to adapt its extraction behavior to the specific characteristics of each domain, resolving the contradiction between accuracy on edited documents and adaptability to live transcripts.
Solution Approach 2:
The patent segments the extraction system into multiple domain-specific models (scientific article extractor and live transcript extractor), each trained on and optimized for its specific domain. The system then routes queries to the appropriate domain model based on the input type, allowing high accuracy within each domain while maintaining overall system versatility across multiple domains.
2Productivity
If machine learning systems extract key-phrases from multiple paragraphs independently, then they process each paragraph individually, but they extract redundant key-phrases across paragraphs
Solution Approach 1:
The patent implements feedback mechanisms where the system tracks previously extracted key-phrases across paragraphs and uses this information to guide subsequent extractions. The uniqueness reward function provides feedback that penalizes redundant extractions, encouraging the system to select diverse, non-repetitive key-phrases while maintaining processing efficiency.
Solution Approach 2:
The system dynamically adjusts its extraction behavior based on the state of previously extracted phrases. Rather than treating each paragraph independently with static parameters, the system modifies its extraction strategy in real-time based on accumulated context, balancing processing speed with the need to avoid redundancy through adaptive decision-making.
3Measurement precision
If existing systems are trained in scientific or weblog domains, then they achieve domain-specific performance, but they lack knowledge of characteristics associated with the live-transcript domain
Solution Approach 1:
The patent creates a universal framework that handles multiple domains (scientific articles, weblogs, and live transcripts) through a common architecture. The domain discriminator network automatically detects the input domain and routes it to the appropriate specialized extractor, providing both domain-specific precision and cross-domain versatility within a single system.
Solution Approach 2:
The domain discriminator network acts as an intermediary between the input and the extraction models. It analyzes the input characteristics and determines the appropriate domain, then directs the query to the corresponding specialized extractor. This intermediary layer enables the system to maintain high precision in each domain while achieving broad cross-domain applicability.
Data Source
AI summary
Systems and methods for key-phrase extraction are described. The systems and methods include receiving a transcript including a text paragraph and generating key-phrase data for the text paragraph using a key-phrase extraction network. The key-phrase extraction network is trained to identify domain-relevant key-phrase data based on domain data obtained using a domain discriminator network. The systems and methods further include generating meta-data for the transcript based on the key-phrase data.