Neural Transduction of Letter Position Dyslexia using an Anagram Matrix Representation

Avi BleiweissACL 2020

TL;DR

This paper introduces a novel neural network approach to correct "letter position dyslexia" (LPD), a reading disorder where individuals misplace letters within words. By representing words as "anagram matrices" and using a two-dimensional context window, their model significantly improves the accuracy of restoring jumbled words, offering a promising avenue for real-time reading assistance.

What Problem Does This Paper Address?

Letter Position Dyslexia (LPD) is a specific type of dyslexia affecting 5-17% of school-aged children, characterized by a deficit in encoding letter order within words. This leads to errors like transposing letters, creating non-words (e.g., "slime" becoming "silme") or even real words with different meanings (e.g., "slime" becoming "smile"). While previous research has focused on offline classification of dyslexia types, this paper aims to develop an in-line remediation framework, meaning it can correct errors as they occur. The challenge lies in accurately identifying and correcting these letter transpositions, especially when the resulting jumbled word is semantically unrelated to the original, and doing so efficiently for practical application.

Methods

The core of the proposed solution is a neural network model that treats LPD correction as a sequence transduction problem, similar to machine translation. The key innovations are:

  • Anagram Matrix Representation: Instead of treating words as simple sequences of letters, the authors represent each word as an "anagram matrix." This matrix captures potential letter transpositions by anchoring the first and last letters and permuting the inner ones. This allows the model to consider various scrambled versions of a word.
  • Two-Dimensional Context Window: Traditional word embedding models use a one-dimensional context window. This paper expands this to a two-dimensional context window, leveraging the anagram matrix. This allows the model to learn semantic similarities not only between adjacent words but also between words that are transpositions of each other, even if they appear in different positions within the matrix.
  • Transformer Architecture: The model utilizes a Transformer architecture, known for its effectiveness in sequence-to-sequence tasks like machine translation. This architecture relies on attention mechanisms, allowing it to weigh the importance of different parts of the input sequence when generating the output.

Datasets

  • Shakespeare's Sonnets: The primary dataset used for training and evaluation is a collection of Shakespeare's Sonnets. This dataset was chosen for its rich language and the presence of apostrophes, which can lead to letter omissions.
  • The raw dataset contained 2,154 verses.
  • A cleanup step removed words with eight or more letters.
  • The augmented dataset used for training comprised 130,021 text sequences.
  • Curated LPD Screening Tests: A smaller, hand-curated corpus of 196 English words prone to reading errors was assembled for baseline evaluation.

Key Insights and Ideas

  • The power of 2D context: The most significant insight is that a two-dimensional context window, when applied to the anagram matrix representation, dramatically improves the model's ability to capture semantic similarities between transposed words. This suggests that understanding the relationships between different permutations of letters is crucial for LPD remediation.
  • Anagram matrices as a rich representation: Representing words as anagram matrices provides a structured way to encode potential letter position errors, making them more amenable to neural network processing.
  • Transformer's suitability for LPD: The Transformer architecture, with its attention mechanisms, proves effective in handling the complex relationships within the anagram matrices and generating coherent, corrected text sequences.

Key Results

  • BLEU Scores: The model achieved a unigram BLEU score of 40.6 on the augmented Sonnets dataset when using a two-dimensional context window. This is a significant improvement over the 36.8 achieved with a one-dimensional window. BLEU scores decreased for higher n-grams (BLEU-4 was 8.9), which is typical.
  • SARI Scores: The SARI (System Alignment and Rewrite Index) metric, which is better suited for text simplification tasks, showed a score of 23.7 for the two-dimensional context window, compared to 21.2 for the one-dimensional. The authors note that SARI scores correlate well with BLEU but favor models that simplify input verses.
  • Screening Test Performance: On the smaller, curated screening corpus, the model achieved a low 1-gram BLEU score of 9.2, which the authors attribute to the dataset's small size and the transformer's tendency to overfit.

Limitations and Future Work

  • Dataset Size: The authors acknowledge that the screening corpus was too small for effective training, leading to overfitting. They emphasize the need for larger datasets that provide sufficient context for learning embeddings.
  • Context Fragmentation: The standard Transformer's fixed-length context can lead to issues with long sequences. The authors suggest exploring architectures like Transformer-XL, which can handle longer dependencies.
  • Real-world Application: Future work includes using the model in real-world LPD restoration scenarios and exploring its efficacy in non-language modeling tasks. They also envision pre-training the model on multiple corpora and developing a responsive system for individuals to compare corrected text side-by-side.

Summary updated

BibTeX
@inproceedings{bleiweiss2020neural,
  author = {Avi Bleiweiss},
  title = {{Neural Transduction of Letter Position Dyslexia using an Anagram Matrix Representation}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2020},
}
Figure