TL;DR
This paper introduces and evaluates a French text-to-pictograph translation system designed to improve communication for individuals with intellectual disabilities in medical settings. By analyzing speech therapist strategies and evaluating the system's "medical coverage," the authors identify areas for improvement to make medical dialogues more accessible.
What Problem Does This Paper Address?
Effective communication between healthcare providers and patients is crucial, but it can be a significant challenge for individuals with intellectual disabilities (ID). These individuals often rely on Augmentative and Alternative Communication (AAC) systems, which can include pictographs, to express themselves. While existing AAC systems and translation tools exist, there's a need for specialized systems that can accurately translate medical language into pictographs, bridging the gap in understanding between doctors and patients with ID. This paper specifically addresses the lack of a robust French text-to-pictograph system for medical contexts and investigates how to best achieve this translation.
Methods
The authors present a French text-to-pictograph translation system that builds upon previous work. The system utilizes a shallow linguistic analysis pipeline (sentence detection, tokenization, POS-tagging, lemmatization) and then translates words through either a semantic route (using the WOLF database and its relations like hyperonyms and antonyms) or a direct route (using a custom dictionary of pictograph filenames). Key improvements to the system include an updated pictograph database with more medical symbols, the implementation of a color-coding system based on POS tags (inspired by Fitzgerald, 1949), and the addition of rules for compressing French sentence structures to simplify them for patient instructions.
To understand how to best translate medical sentences into pictographs, the researchers collaborated with a retired speech therapist. They manually translated 100 medical sentences into Arasaac pictographs, identifying at least 10 distinct translation strategies. These strategies include deletion, insertion of paraphrases, moving words, using synonyms, hyperonyms, hyponyms, antonyms, changing POS tags, and replacing multi-word expressions with single words or other word replacements.
Datasets
- Arasaac pictograph set: An open-source database of over 12,000 pictographs, used as the target output for the translation system. The paper notes that as of November 2021, it contained 1,126 medical pictographs.
- BabelDr medical sentences: 150 French medical sentences from the BabelDr system were used for system tuning and evaluation. These sentences included physician questions, patient instructions, and greetings.
- Manual translations by speech therapist: 100 sentences from the BabelDr dataset were manually translated into Arasaac pictographs by a speech therapist to identify translation strategies.
- 700 sentence transcripts and canonicals: Used for manual evaluation of untranslated words.
Key Insights and Ideas
- Pictograph translation is more than literal substitution: The manual translation experiment revealed that a simple word-for-word lookup is insufficient. Sophisticated strategies involving paraphrasing, synonymy, and structural changes are necessary to convey meaning effectively.
- The importance of context and user needs: The system's development is informed by the need to simplify medical language for individuals with ID, leading to the inclusion of compression rules for patient instructions.
- Arasaac is a valuable resource, but not exhaustive: While Arasaac provides a substantial number of pictographs, the study highlights the ongoing need to expand its medical domain coverage and the system's ability to handle terms not directly represented.
- POS-tagging and semantic relations are key: The system's reliance on POS-tagging and semantic relations (like hyperonyms) from WOLF is crucial for finding appropriate pictograph representations when direct matches are unavailable.
Key Results
- Automated evaluation: The system achieved a BLEU score of 30.3 (±2.3) and WER of 55.5 (±2.1) on medical data, comparable to previous studies. Adding the "xpos relation" (which leverages synonymy and hyperonymy) did not significantly improve BLEU but reduced untranslated tokens.
- Untranslated words: The system left 126 types (191 tokens) untranslated in transcripts and 102 types (229 tokens) in canonicals when not using the "xpos relation." With the "xpos relation," these numbers dropped to 110 types (159 tokens) and 81 types (203 tokens), respectively.
- Lexical coverage: Manual evaluation of 50 canonicals showed that the system achieved a medical coverage of 62.1% (±75.4%) for annotator 1 and 71.8% (±82.8%) for annotator 2, with synonymy being the most frequently used relation.
- Common untranslated words: Adjectives (e.g., "régulier" - regular) and nouns (e.g., "type" - type) were identified as more difficult to translate into pictographs.
Limitations and Future Work
- Limited medical pictograph availability: The study acknowledges the scarcity of medical pictographs in existing sets, particularly for specific French medical terms.
- Word sense disambiguation: The current French system does not handle word sense disambiguation, which can lead to incorrect translations when a word has multiple meanings.
- Sentence comprehensibility: While the system translates text into pictographs, the comprehensibility of these pictographs by target users (individuals with ID) needs further investigation.
- Future work suggestions: The authors suggest improving the system by specializing it for medical dialogues, incorporating word sense disambiguation, exploring other lexical resources and NLP techniques, and conducting further user studies to evaluate sentence comprehensibility.
Summary updated