Error Profiling: Toward a Model of English Acquisition for Deaf Learners

Lisa N. Michaud, Kathleen F. McCoyACL 2001 · DOI

Introduction

For many deaf individuals in the United States, American Sign Language (ASL) is their primary language (L1), while English is acquired as a second language (L2). Because ASL and English possess fundamentally different grammatical structures, deaf learners often face distinct challenges when writing in English. Researchers are currently developing a statistical "error profile" to map this acquisition process, aiming to build a user model for an intelligent tutoring system that can predict a student's "learning frontier" and provide more relevant, personalized feedback.

The Disambiguation Problem in ICICLE

The core of this research supports the development of ICICLE (Interactive Computer Identification and Correction of Language Errors), an AI-driven tutoring system designed specifically for deaf learners. A major hurdle for such systems is the "disambiguation" problem: a single ungrammatical sentence can often be interpreted in multiple ways depending on the student's underlying intent and proficiency level.

Figure 1: Different interpretations of the same error based on learner proficiency level.

As shown in Figure 1, the sentence "She is teach piano on Tuesdays" could be interpreted as a simple morphological error for a beginner, a missing suffix for an intermediate learner, or a failed passive construction for an advanced student. To provide the most relevant pedagogical response, the system must maintain an accurate User Model that predicts the student's current proficiency.

Figure 2: The ICICLE system architecture, highlighting the interaction between the User Model and the Error Identification/Response Generation modules.

Methodology

To establish a predictable ordering of grammatical acquisition, the researchers analyzed a specialized corpus through the following steps:

  1. Corpus Annotation: Writing samples from 106 deaf individuals were tagged with 47 distinct error codes, such as adjective placement and missing determiners.
  2. Expert Benchmarking: Samples were scored using the national Test of Written English (TWE) scale (1–6) by professional ESL instructors.
  3. Statistical Analysis: Using Multivariate Analysis of Variance (MANOVA), the team grouped samples into low, middle, and high proficiency levels to determine if proficiency could statistically predict specific error frequencies while controlling for sample length.

The Progression of English Acquisition

The study reveals that English acquisition for deaf learners follows a visible progression rather than a random accumulation of skills. This progression is often framed within the Zone of Proximal Development (ZPD)—the "frontier" where a student is currently mastering new features. A tutoring system should target this frontier rather than correcting simple slips or advanced structures far beyond the student's reach.

Proficiency Signatures

  • Low Proficiency: Characterized by high rates of "no parse" results (where the system's grammar rules cannot make sense of the sentence) and errors in basic morphology, such as plural markers and simple word order.
  • Mid Proficiency: Learners at this stage often exhibit "ASLisms"—errors where English constituents (like the verb "to be" or determiners) are omitted because they are not explicitly used in ASL grammar. They may also drop subjects or objects that were already established in the discourse context.
  • High Proficiency: These learners begin attempting complex constructions like the passive voice. While their total error count is lower, their mistakes often stem from the "over-generalization" of sophisticated rules they are just beginning to grasp.

Conclusion and Future Work

While the current model effectively identifies what students get wrong, the authors acknowledge that a complete model of acquisition must also track what they get right. Future work involves automating the detection of successfully executed grammatical rules to allow the SLALOM user model to project probabilities of mastery across the entire English grammar. By targeting the "learning frontier," ICICLE can move beyond simple error correction toward truly personalized language instruction.

Summary updated

BibTeX
@inproceedings{michaud2001error,
  author = {Lisa N. Michaud and Kathleen F. McCoy},
  title = {{Error Profiling: Toward a Model of English Acquisition for Deaf Learners}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2001},
  doi = {10.3115/1073012.1073063},
}
Figure