A Blissymbolics Translation System

Usman Sohail, David TraumNAACL 2019

TL;DR

This paper introduces a prototype system for translating Blissymbolics, a pictographic language used by individuals with communication disorders, into natural English. By focusing on morphological realization and employing an N-gram language model, the system aims to reduce reliance on human interpreters and enhance the expressiveness of Blissymbolics communication.

What Problem Does This Paper Address?

Communication disorders affect a significant portion of the population, and for those who use Blissymbolics, a pictographic writing system designed for easier comprehension than alphabetic systems, there's a substantial reliance on human interpreters to bridge the gap to the general population. While Blissymbolics offers an alternative for individuals with low literacy or speech impairments, current tools are limited. Existing systems often focus on creating customized Bliss charts, which restrict expressiveness, or offer limited translation capabilities that don't allow for the creation of new words or conjugations. This paper tackles the challenge of developing a more robust and expressive translation system from Blissymbolics to English, aiming to empower users and decrease their dependence on intermediaries.

Methods

The authors developed a translation system in Python, making several key assumptions about Blissymbolics usage. Firstly, they assumed input sequences would adhere to word-separating tokens as per Blissymbolics' fundamental rules. Secondly, they utilized a 4-5 digit numeric ID encoding scheme from the Official Blissymbolics Dictionary, accepting these IDs as input. The core of their work lies in two main components:

  • Morphological Realization: This component aims to enable the expression of words or conjugations not explicitly present in the official dictionary. They leveraged the SimpleNLG realizer to handle recognized Bliss words and apply morphological transformations, such as past tense realization. This allows for the creation of new word forms based on existing ones.
  • Language Model: To select the most probable English translation from a set of possibilities and to decide on the inclusion of articles, an N-gram language model was implemented. This model was trained on a combination of corpora (Gutenberg, brown, conll2000, and nps-chat) using interpolation smoothing.

Datasets

The paper mentions the use of the Official Blissymbolics Dictionary for character and word IDs. For training the language model, the authors used the Gutenberg, brown, conll2000, and nps-chat corpora. The system was evaluated on a custom test set comprising 15 Bliss utterances sourced from children's books.

Key Insights and Ideas

  • Morphological Realization is Crucial for Expressiveness: The paper highlights that Blissymbolics' strength lies in its ability to form new words and conjugations. A translation system that can handle these variations, even if not officially codified, is essential for true expressiveness and mirrors the flexibility intended by the Blissymbolics community.
  • Encoding is a Major Hurdle: The lack of a standardized and comprehensive encoding scheme for Blissymbolics, especially for modified symbols, significantly complicates the development of computer-assisted tools. This limitation directly impacts the system's ability to differentiate between similar symbols.
  • Context is King (and Currently Lacking): The current system struggles with context-dependent translations, as seen in examples where names like "Julius" are not inferred. A system that can exploit dialogue context would be a significant step towards more natural and accurate translations.
  • Bridging the Gap for Low Literacy Users: By aiming to translate Blissymbolics to English, the system directly addresses the needs of individuals with low literacy, offering them a more accessible way to communicate with the wider world.

Key Results

The translation system achieved a BLEU score of 34.53 on the 15-utterance test set. While some sentences were translated accurately, others contained errors. The paper points out specific error types:

  • Language model errors, such as missing articles.
  • Blissymbolics-related errors, including misinterpretations of symbols and issues arising from the encoding scheme (e.g., differentiating "other" from "you" due to shared IDs for modified symbols).
  • The system successfully demonstrated the ability to translate recognized Bliss words with morphological variations, such as adding a past tense indicator to "cry" to form "cried."

Limitations and Future Work

The authors acknowledge several limitations and outline promising directions for future work:

  • Encoding Scheme: A more robust and computer-friendly encoding scheme is needed that can capture all Blissymbolics capabilities, including symbol modifications.
  • Language Model: The current language model is a proof of concept. Future work should involve training on dialogue-centric corpora to better suit the intended use of the system.
  • User Interface: A critical component for usability is a graphical user interface that allows users to easily input Bliss characters. This also necessitates the selection of an encoding scheme first. A text-to-speech component is also crucial for user independence.
  • Contextual Understanding: Developing a system that can exploit dialogue context is vital for reducing reliance on human interpreters and improving translation accuracy.
  • Morphological Realization Scope: The current system is limited to morphological realization supported by SimpleNLG. Blissymbolics has over 40 morphological relationships, many of which require custom realization mechanisms.
  • New Word Creation: The system currently only applies morphological realization to recognized Bliss words. Expanding this to handle the creation of entirely new words based on Blissymbolics' fundamental rules is a future goal.

Summary updated

BibTeX
@inproceedings{sohail2019blissymbolics,
  author = {Usman Sohail and David Traum},
  title = {{A Blissymbolics Translation System}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2019},
}
Figure