Home › Categories › Technology› BanditMTL: Bandit-based Multi-task Learning For Text Classification

BanditMTL: Bandit-based Multi-task Learning For Text Classification.pdf

2021.acl-long.428.pdf
Preview of BanditMTL: Bandit-based Multi-task Learning for Text Classification
🔗 Source: aclanthology.org
📊 Size: 1.5 MB
👤 Author: Yuren Mao ; Zekai Wang ; Weiwei Liu ; Xuemin Lin ; Wenbin Hu
⬇️ Downloads: 54

Summary

This addresses a critical issue in MTL: when tasks compete, uncontrolled task variance can lead to overfitting in some tasks and underfitting in others, degrading overall performance.

Key Contributions:

1. Task Variance Regularization: The authors highlight the importance of controlling task variance during MTL training. Existing methods ignore this aspect, potentially hindering generalization.

2. BanditMTL Algorithm: BanditMTL incorporates a mirror gradient ascent-descent algorithm within a linear adversarial multi-armed bandit framework. This allows it to jointly minimize empirical losses and regularize task variance.

3. Experimental Validation: Extensive experiments on sentiment analysis and topic classification demonstrate that BanditMTL significantly outperforms existing state-of-the-art MTL methods, validating its effectiveness in improving generalization across tasks.

Methodology:

Multi-Task Learning Formulation: The model jointly optimizes multiple task-specific loss functions using a shared set of parameters.

Adversarial Multi-Armed Bandit Integration:

Treat each task as a "arm" in a bandit algorithm.
 Use a mirror gradient ascent-descent algorithm to balance exploration (trying different tasks) and exploitation (focusing on promising tasks).
Regularizes the variance between task losses, preventing excessive competition or neglect of certain tasks.

Linear Combination Strategy: The model optimizes a weighted sum of individual task losses to achieve balanced performance across all tasks.

Results:

BanditMTL demonstrates superior performance compared to baseline MTL methods and state-of-the-art multi-task text classification approaches on benchmark datasets, highlighting the benefits of controlling task variance.

Download File

Description

By effectively regularizing tasks, BanditMTL aims to improve the generalization of MTL models.

Technical Information

  • File Format: PDF
  • File Size: 1.5 MB
  • Pages: 11
  • Language: EN
  • Author: Yuren Mao ; Zekai Wang ; Weiwei Liu ; Xuemin Lin ; Wenbin Hu
  • Total Downloads: 54
  • Last Updated: 2 hours ago

Document Overview

This PDF document about BanditMTL: Bandit-based Multi-task Learning for Text Classification provides comprehensive information and guidance. Whether you're a beginner or advanced user, this resource offers valuable insights into BanditMTL: Bandit-based Multi-task Learning for Text Classification.

Related Topics

If you're interested in BanditMTL: Bandit-based Multi-task Learning for Text Classification, you might also want to explore:

BanditMTL: guideBanditMTL: tutorialBandit-based guideBandit-based tutorialMulti-task guideMulti-task tutorialLearning guideLearning tutorialText guideText tutorialClassification guideClassification tutorial

Related Categories

Business & Finance
41204 documents
Education & Learning
22711 documents
Health & Wellness
8956 documents

Related Documents

Preview of classification task
PDF file

classification task

95 KBAidan SammonsEN1 page
1. Max: Procedural (skateboard trick) and Episodic (remembering the first successful performance) 2.
Download
Preview of Apathy Classification
PDF file

Apathy Classification

641 KBS L Happy, Antitza Dantcheva, Abhijit Das, Francois F Bremond, Radia Zeghari, Philippe RobertEN7 pages
Apathy classification is proposed using a multi-task learning framework based on facial analysis, leveraging emotion and facial movements, and...
Download
Preview of ACL Proceedings
PDF file

ACL Proceedings

639 KBAssociation for Computational LinguisticsEN10 pages
Researchers propose an adversarial multi-task learning framework for text classification, which separates shared and private feature spaces using...
Download
Preview of Home Learning Task
PDF file

Home Learning Task

290 KBJessica BriggsEN2 pages
The text is a homework assignment with 11 questions about a chapter, including vocabulary, plot, and character analysis, and a drawing task based on...
Download
Preview of Adversarial Text Classification
PDF file

Adversarial Text Classification

660 KBEN19 pages
Researchers propose RSMI, a two-stage framework combining randomized smoothing (RS) and masked inference (MI) to improve adversarial robustness of...
Download

Download BanditMTL: Bandit-based Multi-task Learning for Text Classification eBooks for free and learn more about BanditMTL: Bandit-based Multi-task Learning for Text Classification. These books contain exercises and tutorials to improve your practical skills, at all levels!

Not satisfied with this document? We have related documents to BanditMTL: Bandit-based Multi-task Learning for Text Classification, try searching with similar keywords: BanditMTL: Bandit-based Multi-task Learning for Text Classification, "Hinder och möjliggörare för 1.5°-livsstilar: Ytliga och djupgående strukturella faktorer som påverkar potentialen för hållbar k, Ändring av genomföranderam för en europeisk plattform för utbyte av balansenergi från frekvensåterställn ingsreserver med manuell, Rekommendationer för vaccination mot covid-19 för särskilda grupper av barn -, förstudie för att utvärdera förutsättningarna att genom en innovationsupphandli ng utveckla en drifttjänst för geoenergilager, Självkänsla och KBT ‐ Påverkas självkänslan vid KBT för depression och ångesttillstånd?Se lf‐esteem and CBT ‐ How does CBT for de, Matglädje för alla: en guide till rätt konsistens för olika behov, Task Based Learning And Learning Circles Ppt

You can download PDF versions of the user's guide, manuals and ebooks about BanditMTL: Bandit-based Multi-task Learning for Text Classification, you can also find and download for free A free online manual (notices) with beginner and intermediate, Downloads Documentation, You can download PDF files (or DOC and PPT) about BanditMTL: Bandit-based Multi-task Learning for Text Classification for free, but please respect copyrighted ebooks.