Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Disambiguation of morpho-syntactic features of African American English -- the case of habitual be

Harrison Santiago, Joshua Martin, Sarah Moeller, Kevin Tang

TL;DR

The paper tackles NLP bias against African American English by addressing the rare and ambiguous habitual 'be' in AAE. It introduces a two-stage pipeline—rule-based syntactic filters to prune non-habitual instances and data augmentation to balance habitual ones—enabling unbiased classifiers trained on CORAAL data. On a corpus with about $477$ habitual and $4{,}656$ non-habitual cases, the ensemble classifier reaches approximately $F_1=0.65$, with reduced variance when using the balanced data. This approach demonstrates a practical strategy to mitigate low-resource, dialect-specific features in NLP and could extend to other AAE morphosyntactic phenomena and future transformer-based models as data availability increases.

Abstract

Recent research has highlighted that natural language processing (NLP) systems exhibit a bias against African American speakers. The bias errors are often caused by poor representation of linguistic features unique to African American English (AAE), due to the relatively low probability of occurrence of many such features in training data. We present a workflow to overcome such bias in the case of habitual "be". Habitual "be" is isomorphic, and therefore ambiguous, with other forms of "be" found in both AAE and other varieties of English. This creates a clear challenge for bias in NLP technologies. To overcome the scarcity, we employ a combination of rule-based filters and data augmentation that generate a corpus balanced between habitual and non-habitual instances. With this balanced corpus, we train unbiased machine learning classifiers, as demonstrated on a corpus of AAE transcribed texts, achieving .65 F$_1$ score disambiguating habitual "be".

Disambiguation of morpho-syntactic features of African American English -- the case of habitual be

TL;DR

The paper tackles NLP bias against African American English by addressing the rare and ambiguous habitual 'be' in AAE. It introduces a two-stage pipeline—rule-based syntactic filters to prune non-habitual instances and data augmentation to balance habitual ones—enabling unbiased classifiers trained on CORAAL data. On a corpus with about habitual and non-habitual cases, the ensemble classifier reaches approximately , with reduced variance when using the balanced data. This approach demonstrates a practical strategy to mitigate low-resource, dialect-specific features in NLP and could extend to other AAE morphosyntactic phenomena and future transformer-based models as data availability increases.

Abstract

Recent research has highlighted that natural language processing (NLP) systems exhibit a bias against African American speakers. The bias errors are often caused by poor representation of linguistic features unique to African American English (AAE), due to the relatively low probability of occurrence of many such features in training data. We present a workflow to overcome such bias in the case of habitual "be". Habitual "be" is isomorphic, and therefore ambiguous, with other forms of "be" found in both AAE and other varieties of English. This creates a clear challenge for bias in NLP technologies. To overcome the scarcity, we employ a combination of rule-based filters and data augmentation that generate a corpus balanced between habitual and non-habitual instances. With this balanced corpus, we train unbiased machine learning classifiers, as demonstrated on a corpus of AAE transcribed texts, achieving .65 F score disambiguating habitual "be".
Paper Structure (15 sections, 1 figure, 2 tables)

This paper contains 15 sections, 1 figure, 2 tables.

Table of Contents

  1. Introduction
  2. Related work
  3. Habitual "be"
  4. Methodology
  5. Preprocessing
  6. Rule-based filter of non-habitual "be"
  7. Augmenting habitual "be"
  8. Classifiers
  9. Experiment
  10. Data
  11. Results
  12. Conclusion
  13. Appendix: Types of non-habitual "be"
  14. Rules to filter non-habitual "be"
  15. Examples of augmenting occurrences of the habitual "be"

Figures (1)

  • Figure 1: The disambiguation pipeline: the input corpus goes through a Part-of-Speech tagger, after which non-habitual instances are separated by a rule-based filter. Any indeterminate "be" instances are balanced by augmentation and tagged by classification models.