Datasets:
Edit json files with BCP-47 code as 'en'
Question about gloss annotations in the ISL Sign Dictionary dataset
Hello,
I am developing a bidirectional Indian Sign Language translation system and am evaluating your sign dictionary as a potential vocabulary resource.
The README describes 3,077 lexical items and demonstrates metadata fields named transcripts and glosses. However, the dataset viewer currently exposes a transcript object with an English text field in the sample records, and I could not find a separate glosses field.
Could you please clarify:
- Is the English transcript itself intended to serve as the gloss label for each sign video?
- Does the actual dataset contain a separate ISL gloss annotation that is not exposed in the viewer?
- Are there documented conventions for gloss labels, synonyms, compounds and ambiguous English terms?
- Is the dataset intended primarily for lexical sign recognition, or can it also support sentence-level translation research?
I will retain attribution and comply with the CC BY-SA 4.0 license.
Thank you for making this resource available.
Hello,
Thank you for your interest in the ISL Sign Dictionary, and for taking the time to look through the dataset so carefully. It's great to hear it may be useful for your bidirectional ISL translation work. Here are answers to your questions.
1. Is the English transcript intended to serve as the gloss label?
Yes. Each video in this dataset contains a single sign, so the English word in the text field of the transcript object works as the gloss label for that sign (for example, "text": "rehoboam"). For this dataset, "transcript" and "gloss" mean the same thing.
2. Is there a separate ISL gloss annotation not shown in the viewer?
No, there's no hidden gloss field in this dataset. The distinction between transcript and gloss matters more in our companion dataset, ISL Sign Bibles. There, each video contains longer discourse: paragraphs of several sentences or Bible verses. The transcript in that dataset is the English equivalent of the signed content.
When we use the term gloss, we mean a single sign word. A gloss annotation for a longer video would list the signs in the order they appear. Because ISL and English differ in vocabulary and word order, that sequence reads quite differently from the English transcript. These gloss annotations aren't available yet. We're currently annotating a subset of the Sign Bibles dataset and hope to release them in the future.
3. Are there documented conventions for glosses, synonyms, compounds and ambiguous terms?
Not at the moment. We don't have formal documented conventions for gloss labels, synonyms, compounds or ambiguous English terms. Each English label simply identifies the sign shown in the video, so we'd suggest treating these labels as identifiers rather than as standardised glosses.
That said, the labels can still be a useful starting point for handling synonyms and ambiguity in your system. One approach is to combine the English words from our data with external English language resources. For example, a thesaurus or a lexical database such as WordNet could map synonyms or related words to an existing sign, and word embeddings could help find the closest available sign for an English word that isn't in the dictionary. Since many English words have several meanings, a context-aware method, such as contextual embeddings, may help choose the right sign for ambiguous terms during English-to-ISL translation.
4. Lexical recognition or sentence-level translation?
The Sign Dictionary is mainly suited to lexical, isolated-sign tasks such as sign recognition and vocabulary building. For sentence-level or discourse-level translation research, the Sign Bibles dataset is a better fit, since it pairs continuous signing with English transcripts. Used together, the two datasets may support both directions of your system.
Thank you for your commitment to attribution and to the CC BY-SA 4.0 license. We really appreciate it. Please feel free to reach out with any further questions, and we'd love to hear how your project develops.
Best regards