A Multi-Context Character Prediction Model for a Brain-Computer Interface

Shiran Dudy, Shaobin Xu, Steven Bedrick, David SmithNAACL 2018

TL;DR

This paper introduces the Online-Context Language Model (OCLM) for Brain-Computer Interfaces (BCIs). OCLM improves character prediction accuracy and speed by integrating both character-level and word-level language models, and by handling ambiguous user input, which is crucial for noisy BCI signals.

What Problem Does This Paper Address?

Brain-Computer Interfaces (BCIs) offer a lifeline for individuals with severe communication disabilities, enabling them to interact with their environment. However, current BCI systems, especially those relying on electroencephalography (EEG), face significant challenges. EEG signals are inherently noisy, making it difficult to accurately infer user intent. This leads to slow communication rates and a high potential for errors, causing user frustration. Traditional methods often struggle with ambiguous input and out-of-vocabulary (OOV) words, further hindering usability. This paper aims to address these limitations by developing a more robust and adaptive character prediction model for BCIs.

Methods

The core innovation is the Online-Context Language Model (OCLM). OCLM addresses the ambiguity of BCI input by maintaining multiple possible interpretations of the user's intent at each step. It achieves this by:

  • Multi-Contextual Input: OCLM splits the input history into two parts: the "word history" (completed words) and the "trailing prefix" (the current word being typed).
  • Word-Level Language Model (FwLM): It uses a word-level language model to predict likely word completions based on the trailing prefix. This helps in generating possible word candidates.
  • Character-Level Language Model (FcharLM): It also incorporates a character-level language model to predict the next character.
  • Fusion and Ambiguity Handling: OCLM fuses the word and character predictions to create a weighted distribution of likely next characters. Crucially, it doesn't commit to a single prediction but maintains multiple candidates, allowing for re-evaluation as more data becomes available. This is implemented using Finite State Transducer (FST) operations.
  • Out-of-Vocabulary (OOV) Word Support: The model is designed to handle words not present in its lexicon by mapping infrequent words to an <unk> token and allowing for partial prefix completion.
  • Auto-Correction: While not explicitly performing character deletion/insertion correction, OCLM's continuous re-evaluation of word history based on new EEG evidence implicitly aids in correcting misinterpretations.

Datasets

The models were trained on the Brown Corpus (80% for training) and evaluated on a test set comprising 10,265 sentences and 176,280 words (20% of the data). For evaluating performance with shorter utterances, a subset of the test set containing 2,954 sentences with 16,519 words (sentences of 10 words or less) was also used.

Key Insights and Ideas

  • Ambiguity is a Feature, Not a Bug: Instead of trying to resolve ambiguity immediately, OCLM embraces it by maintaining multiple hypotheses. This is critical for noisy BCI signals where a single "correct" interpretation might be elusive.
  • Word-Level Context is Crucial: Integrating word-level language models significantly improves character prediction, especially when dealing with in-progress words and OOV words.
  • Early Prediction Matters: The OCLM demonstrates higher performance (Reciprocal Rank) earlier in word prediction, which is vital for improving the overall typing speed and user experience in BCIs.
  • Robustness to Noise: OCLM shows more graceful degradation in performance compared to other models when faced with ambiguous input, highlighting its durability.

Key Results

  • Deterministic History (n=1):
    • OCLM and PreLM (a prefix language model) significantly outperformed NGRAM (a smoothed character 5-gram model) in Mean Reciprocal Rank (MRR) and ACC@10.
  • OCLM and PreLM showed similar overall performance, but OCLM had a slight edge in predicting earlier letters within a word.
  • Ambiguous History (n=2, n=3):
    • OCLM maintained a significantly higher MRR and ACC@10 compared to PreLM, even as the ambiguity increased.
  • While OCLM's performance degraded with more ambiguity, PreLM experienced a more substantial breakdown.
  • OCLM's higher MRR persisted for a longer duration within words, especially for n=2.

Limitations and Future Work

  • Limitations:
    • The current auto-correction mechanism is implicit, focusing on re-evaluating word history rather than explicit character deletion/insertion correction.
  • The evaluation was performed using a simulated BCI, not with actual end-users.
  • The paper notes that PreLM fails to run on complete ambiguous history for space reasons.
  • Future Work:
    • Integrate OCLM into a real BCI system and evaluate with end-users to assess its impact on sequence display reduction and typing time.
  • Further investigate OOV word handling and explore user-level personalization.
  • Explore more concrete auto-correction approaches.
  • Investigate potential applications in spelling correction.

Summary updated

BibTeX
@inproceedings{dudy2018multi,
  author = {Shiran Dudy and Shaobin Xu and Steven Bedrick and David Smith},
  title = {{A Multi-Context Character Prediction Model for a Brain-Computer Interface}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2018},
}
Figure