Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

An Analysis under a Unified Fomulation of Learning Algorithms with Output Constraints

Mooho Song, Jay-Yoon Lee

TL;DR

This work examines all the algorithms on three NLP tasks: natural language inference (NLI), synthetic transduction examples (STE), and semantic role labeling and explores and reveals the key factors of various algorithms associated with achieving high $H\beta$-scores.

Abstract

Neural networks (NN) perform well in diverse tasks, but sometimes produce nonsensical results to humans. Most NN models "solely" learn from (input, output) pairs, occasionally conflicting with human knowledge. Many studies indicate injecting human knowledge by reducing output constraints during training can improve model performance and reduce constraint violations. While there have been several attempts to compare different existing algorithms under the same programming framework, nonetheless, there has been no previous work that categorizes learning algorithms with output constraints in a unified manner. Our contributions are as follows: (1) We categorize the previous studies based on three axes: type of constraint loss used (e.g. probabilistic soft logic, REINFORCE), exploration strategy of constraint-violating examples, and integration mechanism of learning signals from main task and constraint. (2) We propose new algorithms to integrate the information of main task and constraint injection, inspired by continual-learning algorithms. (3) Furthermore, we propose the $Hβ$-score as a metric for considering the main task metric and constraint violation simultaneously. To provide a thorough analysis, we examine all the algorithms on three NLP tasks: natural language inference (NLI), synthetic transduction examples (STE), and semantic role labeling (SRL). We explore and reveal the key factors of various algorithms associated with achieving high $Hβ$-scores.

An Analysis under a Unified Fomulation of Learning Algorithms with Output Constraints

TL;DR

This work examines all the algorithms on three NLP tasks: natural language inference (NLI), synthetic transduction examples (STE), and semantic role labeling and explores and reveals the key factors of various algorithms associated with achieving high -scores.

Abstract

Neural networks (NN) perform well in diverse tasks, but sometimes produce nonsensical results to humans. Most NN models "solely" learn from (input, output) pairs, occasionally conflicting with human knowledge. Many studies indicate injecting human knowledge by reducing output constraints during training can improve model performance and reduce constraint violations. While there have been several attempts to compare different existing algorithms under the same programming framework, nonetheless, there has been no previous work that categorizes learning algorithms with output constraints in a unified manner. Our contributions are as follows: (1) We categorize the previous studies based on three axes: type of constraint loss used (e.g. probabilistic soft logic, REINFORCE), exploration strategy of constraint-violating examples, and integration mechanism of learning signals from main task and constraint. (2) We propose new algorithms to integrate the information of main task and constraint injection, inspired by continual-learning algorithms. (3) Furthermore, we propose the -score as a metric for considering the main task metric and constraint violation simultaneously. To provide a thorough analysis, we examine all the algorithms on three NLP tasks: natural language inference (NLI), synthetic transduction examples (STE), and semantic role labeling (SRL). We explore and reveal the key factors of various algorithms associated with achieving high -scores.
Paper Structure (38 sections, 11 equations, 6 figures, 4 tables, 1 algorithm)

This paper contains 38 sections, 11 equations, 6 figures, 4 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Unified formulation of previous work
  3. Type of constraint loss
  4. Probabilistic Soft Logic (PSL)
  5. REINFORCE
  6. Exploration of constraint-violating examples
  7. Sampling
  8. Argmax (Top-1)
  9. Exhaustive
  10. Integration mechanism of learning signals from main task and constraint
  11. Static
  12. Monotone ($\lambda \uparrow$)
  13. Projection
  14. Further exploration on integration of main task and constraint information
  15. Motivation
  16. ...and 23 more sections

Figures (6)

  • Figure 1: Experiment result from NLI task with samp-10. Three bar plots represents the $H\beta$-scores with respect to the integration mechanism (separated by the $x$-axis) and type of constraint losses (separated by the color). From top to bottom, the corresponding values of $\beta$'s are $0.3$, $1$, $3$, respectively.
  • Figure 2: Experiment result from STE task with samp-10. Three bar plots represents the $H\beta$-scores with respect to the integration mechanism (separated by the $x$-axis) and type of constraint losses (separated by the color). From top to bottom, the corresponding values of $\beta$'s are $0.3$, $1$, $3$, respectively.
  • Figure 3: Experiment result from SRL task with samp-10. Three bar plots represents the $H\beta$-scores with respect to the integration mechanism (separated by the $x$-axis) and type of constraint losses (separated by the color). From top to bottom, the corresponding values of $\beta$'s are $0.3$, $1$, $3$, respectively.
  • Figure 4: The $H\beta$-score values for different values of $\beta$ for three tasks: NLI, STE and SRL. For each $\beta$, the top 5 experimental results with the highest $H\beta$-scores are presented for each of the five integration mechanisms described in section §\ref{['subsec:update']}.
  • Figure 5: The $H\beta$-score values for different values of $\beta$ for three tasks: NLI, STE and SRL. For each $\beta$, the top 5 experimental results with the highest $H\beta$-scores are presented for each of the three types of constraint losses described in section §\ref{['subsec:Constraint_Loss']}.
  • ...and 1 more figures