Scripture Forge: Translation Suggestions

Raymond Luong

SIL Language Technology, Scripture Forge Development Team

October 2023

Scripture Forge/NLP Development

Scripture Forge

​

Nathaniel Paulus (Tech lead)

​

Mark Strobert

Raymond Luong

Nigel Wells

Peter Chapman

Colin Moore

Joseph Myers

NLP (Natural Language Processing)

​

Michael Martin (Project lead)

​

David Baines

Damien Daspit

John Lambert

Eli Lowry

Bethany Moore

Matthew Shannon

Tom Albers (Manager)

What is Scripture Forge?

• A browser-based, real-time translation environment for drafting, editing, and community checking of your Paratext project

​

• A home for accessing AI and web innovations available to augment your project

​

• Requires a Paratext project, uses your existing registration

​

• Syncs directly with Paratext

​

• Available online at ScriptureForge.org

What can Scripture Forge do for you?

Online

Browser-based

Integrated with Paratext

Community Checking

​

• Engage the language community in reviewing a draft translation
• Question / answer format
• Easy bulk loading of checking questions
• Paratext Notes integration

Translation Editing

​

• Collaborative online editing environment for Bible translators
• Simplified interface for easier, quicker translator onboarding

Translation Suggestions

​

• Adapt a reference translation into a related language
• Optional suggestions offered as translator is drafting

Translation Drafting

​

• Draft translation of the next book in your project or your back translation
• Add to your Paratext project chapter-by-chapter
• Review and edit before team checking

Scripture Forge - Translation Suggestions

Incremental�Learning

Scripture Forge - Translation Suggestions

• Currently uses statistical machine translation (SMT)
• Limited use cases
• Same word order
• Literal translation
• Related language adaptation
• Working on next generation suggestions engine (soon)
• Neural machine translation (NMT)
• Better quality
• More use cases

Outline

• Setting up the project on SF
• Enable Translation Suggestions
• Settings
• Confidence Threshold
• Editing (Mark chapter 1)
• Edit verses 8 and 12.
• Add text to verse 15.

Translation Suggestions Demo

Machine Drafting: What Factors Support Good Drafts?

​

​

​

​

​

Good results depend on -

• Good Data
• Quality Data: a checked and clean translation will give the best results.
• Sufficient Data: NT is a good starting point, making drafting useful for OT conversion projects.
• Good Alignment Between Source and Target
• An up-to-date back translation in the same language as the source often gives the best results.
• A second good option is a published Bible that has been adapted or used as a single source.
• Good Translation Teams
• Drafts are predicted or “best-guess” translations. They will always contain errors and need to be checked against the source text by an experienced translator.
• Experienced teams can often work with the drafts to produce a quality first draft in less time.

*The list of 200 language in the NLLB model can be found on pp. 13-16 of this paper: https://arxiv.org/ftp/arxiv/papers/2207/2207.04672.pdf

​

THE BASE MODEL:

Drafting is based on the “No Language Left Behind” (NLLB) model by Facebook/Meta, which contains data from 200 world languages.

​

FINE-TUNING:

Translating into a new language

requires parallel sentences between one of the 200 languages and the new language.

​

Machine Drafting

Currently in its early stages

• Based on the “No language left behind” NLLB model (Meta/Facebook)
• Contains data from 200 world languages
• Uses existing translations as input data to “fine tune” the model

​

Coming in early 2024

• Support for Back Translation drafting
• Ideally used for New Testament to Old Testament translations projects

​

Demo of Drafting feature in Scripture Forge

Drafting in the Wild

Drafting in the Wild

• Evaluated MT drafting and quality assessment tools with 7 South Asia translation teams (summer ‘23)
• New-to-Old translation projects
• Focus on early OT books (GEN, RUT, EST, JON, 1SA)
• Sino-Tibetan, Indo-European languages; Devanagari
• Ongoing, growing collaboration

• Onboarded 4 Philippines translation teams
• Mother tongue drafting
• Back translation drafting
• Quality assessments, audio tools for OBT
• Translation suggestions
• Malayo-Polynesian languages; Latin script

Community Checking: Automation, Tracking, Flexibility

​

​

​

​

​

​

​

​

​

​

​

​

​

​

​

​

​

​

Checkers:

• Invite participants via email or text
• Participants respond to questions in mobile app
• Flexible social settings
• Transparent authentication for increased security

​

​

​

​

​

​

​

​

​

​

Translation teams:

• Add questions one-by-one or bulk import
• Questions relate to verse or verse range
• Select answers for export as notes in Paratext
• Audio support for questions and answers

​

​

​

​

​

​

​

​

​

Coming Soon

• Play audio recordings synchronized with the scripture text
• Add comments on any verse

Community Checking Demo in Scripture Forge

What’s Next / Getting in Touch

Fill out the brief form linked here to keep in touch with our team and request follow-up.

​

Explore Scripture Forge at

www.scriptureforge.org

​

For Machine Drafting, please contact

David Baines david_baines@sil.org

Michael Martin michael_martin@sil.org

Bethany Moore bethany_moore@sil.org

​

​

​

Bible Translation Path

1. Exegesis

and Drafting

2. Team check and Comprehension Check

3. Preparing for Consultant

4. Consultant Checking

5. Community Review

6. Final Preparation for Publication

Machine Drafting

Community Checking

Back Translation Drafting