Few-Shot Classification Network for Event Trigger Word Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional event detection systems rely heavily on large amounts of labeled data for each pre-defined event type, making them inefficient for handling unseen event types and insufficient training samples.
Innovation Solution
The proposed system employs a few-shot classification network trained using two sets of labeled samples for each training iteration, where one set is randomly selected and the other is chosen from recently used sets to prevent 'forgetting' and improve scalability.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If conventional supervised learning methods are used for event detection, then the system can achieve accurate classification for pre-defined event types, but the system requires large amounts of labeled data and cannot handle unseen event types
Solution Approach 1:
The system performs preliminary action by storing samples from previous training iterations in a memory buffer before encountering unseen event types. This allows the model to retain knowledge from previously learned event types and apply it to new, unseen types through few-shot learning, reducing the need for large amounts of labeled data for each new event type.
Solution Approach 2:
The system changes the training parameter from requiring large amounts of labeled data per event type to using a small buffer of stored samples from previous iterations. This parameter change enables the model to adapt to unseen event types by leveraging patterns learned from previously seen types, thus improving versatility while reducing data requirements.
2Reliability
If the model is trained with a single randomly selected support set for each iteration, then the training process is simple, but the model suffers from forgetting previously learned event types
Solution Approach 1:
The system ensures continuity of useful action by continuously storing and retaining samples from previous training iterations in a memory buffer. This continuous retention allows the model to maintain knowledge of previously learned event types while learning new ones, preventing catastrophic forgetting without significantly increasing training complexity.
Solution Approach 2:
The system implements feedback by using the stored samples from previous iterations as a reference during current training iterations. The model receives feedback from these stored samples, allowing it to adjust its learning process to maintain performance on previously learned event types while adapting to new ones, thus improving reliability of knowledge retention.
3Measurement precision
If large amounts of labeled data are collected for each event type, then the model can achieve better accuracy, but the system becomes less scalable and more time-consuming to train
Solution Approach 1:
The system uses copying by storing copies of samples from previous training iterations in a memory buffer. Instead of collecting large amounts of new labeled data for each event type, the system copies and reuses previously seen samples, maintaining classification accuracy while significantly improving training efficiency and scalability to unseen event types.
Data Source
AI summary
Systems and methods for natural language processing are described. Embodiments of the present disclosure receive text including an event trigger word indicating an occurrence of an event; classify the event trigger word to obtain an event type using a few-shot classification network, wherein the few-shot classification network is trained by storing first labeled samples during a first training iteration and using the first labeled samples for computing a loss function during a second training iteration that includes a support set with second labeled samples having a same ground-truth label as the first labeled samples; and transmit event detection information including the event trigger word and the event type.