Rule-based Processing and Association-based Processing in Artificial Grammar Learning

Table of Contents

• Introduction
• Requirements

Introduction

Question

In our daily life, different forms of knowledge are involved.

• Rule-based knowledge i.e. When we try to identify whether an integer x is oddd or even, we use a specific rule: x%2==0? Even:Odd
• Association-based knowledge i.e. classical conditioning, operant conditioning

Based on the form of knowledge, two learning and decision making processes are proposed, rule-based learning and association-based learning. In the learning phase, rule-based learning stores abstract rules wherease association-based learning stores associations. When a new case comes, rule-based learning considers whether the new case satisfies a specific rule and then gives a corresponding result. Association-based learning checks whether the new case could trigger stored associations. A judgment depends on the number of triggered associations and the intensity of each association.

The current experiment investigates how a rule-based variable and an association-based variable influence human performance in a simplified experimental setting of learning, the artificial grammar learning paradigm.

ARTIFICIAL GRAMMAR LEARNING PARADIGM (AGL)

• Basic Question

  The artifical grammar learning paradigm is a simplified learning-test experimental setting. After exposure to strings derived from a formal grammar, human participants illustrated above-chance accuracy on grammatical judgments of new strings without knowledge in details of the grammar (Reber, 1967). Strings derived from a formal grammar exhibit not only rule-based patterns but also statistical/associative features. AGL research investigates what type of knowledge, rule-based or association-based or both, has been learned

• Implementation

  Prepare a formal grammar (AG): most researchers used finite state grammars

• Training Phase:

  • Participants will observe/memorize a set of items generated by the AG
  • Participants will not be told that strings follow certain rules

• Test Phase:

  • A list of new items is generated. X% follows the AG (Grammatical-G) and 1-x% does not (Ungrammatical-UG)
  • Participants will be told that training strings follow a certain grammar but not details of the grammar
  • Participants will be asked to determine whether new items are grammatical or ungrammarical

• Transfer Setting

Learning items and test items use the same grammatical rules but different alphabets. i.e. The learning session uses letter sequences and the test session uses color sequences.

• Normal Results:

  • Participants exhibit above chance accuracy
  • Participants could not articulate how they make grammaticality judgments.

• Various Interpretations

  • Rule-Based Interpretations

    • Participants have learned complete or partial of the original grammar (Reber,1967,1969, 1989)

    • Participants have learned a set of propositional rules with the form {Feature -> Grammaticality} (Dulany, et al, 1984)

      Feature refers to a chunk of symbols. Grammaticality refers to "Grammatical" or "Ungrammatical"

      i.e. A participant might learned that items with the chunk "XV" are always grammatical and establishes the rule {"XV" -> Grammatical}

  • Statistics-Based Interpretations

    • Specific Similarity/Edit Distance: Grammatical judgment of a given test item is based on whether the test item is highly similar to a specific learning item (Vokey & Brooks, 1992)
    • Generalized Context Model: Grammatical judgment of a given test item is based on the averaged similarity between the test item and all learning items (Pothos & Bailey, 2000)
    • Analogical Similarity: Grammatical judgment of a transfer test item is based on structural similarity with learning items (Brooks & Vokey)
    • Chunk Strength: Grammatical judgment of a given test item is based on whether the test item contains frequent bigrams or trigrams(Knowlton & Squire, 1996)
    • Entropy: Grammatical judgment of a given test item is based on its entropy value according to all learning items (Jamieson et al., 2016)

PRESENT EXPERIMENT

Independent Variables

1. Grammar Complexity:
   • Finite State grammar
   • Context Free Grammar
2. Chunk Strength: For a given item, chunk strength is the averaged frequency of all its bigrams and trigrams in the learning session

Experiment Design

Both standard and transfer settings are used

2 (Grammatical vs. Ungrammatical) x 3 (High, Medium & Low Chunk Strength) x 2 (Changed Module vs. Unchanged Module)

Requirements

• NLTK
• NumPy