creating meaningful language technology without machine learning

Jonathan N. Washington
Swarthmore College
ClingDing, 2026-03-10

Yukwe ènta kishkwik Lënapehòkink kpëmëtunhalelhumo. Kenahkituneyo Lënapeyòk yu hàki òk Lënapei Sipu lòmëwe nòchi, yukwe, òk apchich. Wëlilìsitàm tìlìch këmaxinkwelëmanëneyo Lënapeyunkahke, òk Lënapeyòk ahpihtit yukwe òk alàpa.

(I am speaking to you today from Lenapehoking, land of the Lenape. We recognise the Lenape as past, present, and future stewards of this land and the Delaware River. Let us honor through our actions the Lenape who were, those who are, and those who will be.)

Overview

My talk in a nutshell

Will overview how to develop meaningful language technology by developing symbolic tools.

The next ~42 minutes:

  • Meaningful language technology
  • Symbolic language technology (LT)
  • Getting started
  • Concluding thoughts

What NLP do we work on?

Common NLP tasks

  • Tokenisation
  • POS Tagging
  • Sentiment analysis
  • Stance detection
  • Inference / textual entailment
  • Semantic role labelling
  • Named entity recognition
  • Speech recognition
  • ...

How we experience them

How real people experience NLP

1
2
3
4

5
6
7
8

Key:

  1. Input methods
    predictive text
  2. Machine translation
  3. Screen readers (TTS)
  4. Spell checking
    grammar checking
  5. Voice assistants
  6. Computer-aided language learning
  7. Search engines
  8. Interfaces (localisation)

Why language technology matters

Why is language technology important to language vitality?

  • Digital technology is hugely pervasive
  • Language mediates most interaction with digital technology

What does language technology do for a language?

  • adds to ease with which a language can be used in digital contexts
  • increases range of uses of language
  • raises perceptions of language
  • supports maintenance and revitalisation efforts

What happens without language technology?

"Digital language death" (i.e., loss of intergenerational transmission)

(Kornai 2013)

Language technology supports language vitality

Reframing NLP

Q: Does this support language vitality?

A: not directly

Meaningful language technology

  • Community-oriented terminology
    • NLP → Language technology (LT)
    • "low-resource" → meaningful in low-vitality situations
  • Evaluation
    • should include (potential) usefulness
    • corporate NLP has this part this right
  • Motivation: how can LT
    • empower communities
    • promote linguistic vitality

What's meaningful
in low-vitality situations?

What's meaningful in low-vitality situations?

not all low-vitality situations are the same (Liu et al. 2022)

Speakers Outside interest Language Technology Resource availability Local expertise in Lg & NLP Community priorities
Language A (350) "small" national language ≥1M, <10M some big corporations some seamless tools; increasing localisation large corpora plenty
  • voice assistants
  • screen readers
  • updates to keyboards that everyone finds annoying
Language B (1K) "large" regional language ≥100K, <1M occasional corporate some tools; spotty localisation copora (spotty) some
  • seamless keyboard instead of stand-alone app
  • localisation
  • search engine support
Language C (3.8K) "small" regional language ≥1K, <100K, increasingly few young some academic a couple local web-based tools some published literature some "activists", mostly self-taught
  • pred. text keyboards
  • localisation
  • spelling and grammar checking
  • MT (for quick media generation)
Language D (764) extremely "endangered" language <100 fluent L1, mostly elders some academic none some published texts, mostly collected narratives adult learners (small community)
  • simple keyboards
  • CALL
  • anything that raises interest & awareness & helps produce new speakers

Symbolic methods

Symbolic tools

  • No (or minimal) machine learning (ML)
    • Requires no large corpora
    • Can be done low-budget, with no special equipment
    • (Essentially) single development cycle (Butt 2020)
    • Highly precise (Butt 2020)
  • Leverage linguistic knowledge
    • Good for partnerships/collaborations with communities
  • Outcomes
    • Practical?
    • Meaningful?
Example: Morphological Transducers

Morphological transducers

Example:eng

Input / Output:isn't
↕
Output / Input: beverbpresp3sg+notadv

Example:kaz

Input / Output:Алмасы
↕
Output / Input: алмаnpx3spnom
алмасnpx3spnom
Алмаnpantfpx3spnom
Алмасnpantmpx3spnom
алvtvneggpr_futsubstpx3spnom
алvauxneggpr_futsubstpx3spnom
алvtvnegger_futpx3spnom
алvauxnegger_futpx3spnom

Morphological transducers

“Downstream tasks” (uses):

  • Spell checkers
  • Component of machine translation (MT) systems
  • Computer-assisted language-learning (CALL)
  • Script conversion
  • Other NLP tasks and research
  • Whatever else you can imagine

Morphological transducers

Development

  • Advantages:
    • No requirement for giant corpora or powerful computers
    • More precise than corpus-based tools (Butt 2020)
    • Easy to fix errors and expand coverage
    • (In MT) lots of "free rides" for related languages
    • Support for multiple orthographies mostly trivial (Washington et al. 2020, Washington et al. 2021)
    • (Essentially) single development cycle (Tyers et al. 2015, Butt 2020)
    • Only need knowledge of language, not mathematics and programming
  • Disadvantages:
    • ...Need knowledge of language, not mathematics and programming
    • Takes time to develop (human time, not compute)
  • i.e., good for community partnerships:
    • Collaboration between different types of experts, as equals
    • Less prone to extractive & other marginalising approaches

Case study: Turkic morphological transducers

Turkic transducers I've contributed to
by naïve coverage & lexicon size + distribution

Production-level
92%-98% coverage
(Tatar, Sakha, Kazakh, Turkish, Kyrgyz, Crimean Tatar, Tuvan)
Working
88-93% coverage
(Chuvash, Uzbek, Bashqort, Qaraqalpaq, Uyghur, Karachay-Balkar, Gagauz, Kumyk)
Prototype
<80% coverage
(Azerbaycani, Iraqi Türkman, Turkmen, Noghay, Khakas, Urum, Altay, Ottoman)
key: not directly involved in, published
(
Washington et al. (2012), Tyers et al. (2012), Salimzyanov et al. (2013), Washington et al. (2014), Tyers et al. (2016), Washington et al. (2016), Bayatlı et al. (2018), Tyers et al. (2019), Gökırmak et al. (2019), Washington et al. (2019), Washington et al. (2020), Ivanova et al. (2021, 2022)
)

Case study: Turkic morphological transducers

Transducers I've contributed to: downstream tasks

Kazakh

  • Used in other NLP research in Kazakhstan
  • Used as segmenter in a shared task
  • Spell checker for Word & LibreOffice
  • Verb paradigm generator

Tatar

  • Used to annotate Tatar National Corpus
  • Spell checker for Word & LibreOffice

Uyghur

  • Used in corpus-based phonology research

Tuvan

  • Used for UniMorph shared task (2021)
  • Spell checker for Word & LibreOffice

Sakha

  • Used in Revita (CALL)
  • Used for UniMorph shared task (2021)

Kyrgyz

  • Spell checker for Word & LibreOffice
  • Used for annotating published corpora

Morphological transducers

Downstream tasks with potential uses by communities

  • Spell checkers
    • Solutions for integration with some environments (Word, LibreOffice)
    • Hard to add at OS level (Windows, macOS, iOS, Android, ChromeOS)
    • Nigh impossible to integrate seamlessly in cloud platforms (Google Docs, MS Office 365)
  • Computer-aided language learning (CALL)
    • Paradigm generators (Apertium Paradigmatrix)
    • Morphological dictionaries (Morphodict, Apertium Dictionary Mode)
    • Text-reading support (Revita, Geriaoueg)
  • Machine Translation (Apertium, Giellatekno)
    • For saving limited translator/editor time when generating content speakers interact with
    • Available as stand-alone websites and local services
    • Possible to integrate into some word processors, browsers

Morphological transducers

Downstream tasks with potential uses by communities

CALL: Apertium (HTML-Tools) Dictionary Mode

Morphological transducers

Downstream tasks with potential uses by communities

MT: Apertium

Case study: Turkic morphological transducers

Turkic transducers I've contributed to
by naïve coverage & lexicon size + distribution

Production-level
92%-98% coverage
(Tatar, Sakha, Kazakh, Turkish, Kyrgyz, Crimean Tatar, Tuvan)
Working
88-93% coverage
(Chuvash, Uzbek, Bashqort, Qaraqalpaq, Uyghur, Karachay-Balkar, Gagauz, Kumyk)
Prototype
<80% coverage
(Azerbaycani, Iraqi Türkman, Turkmen, Noghay, Khakas, Urum, Altay, Ottoman)
key: not directly involved in, published
(
Washington et al. (2012), Tyers et al. (2012), Salimzyanov et al. (2013), Washington et al. (2014), Tyers et al. (2016), Washington et al. (2016), Bayatlı et al. (2018), Tyers et al. (2019), Gökırmak et al. (2019), Washington et al. (2019), Washington et al. (2020), Ivanova et al. (2021, 2022)
)

Case study: Turkic morphological transducers

Transducers I've contributed to: downstream tasks

Kazakh

  • Used in other NLP research in Kazakhstan
  • Used as segmenter in a shared task
  • Spell checker for Word & LibreOffice
  • Verb paradigm generator

Tatar

  • Used to annotate Tatar National Corpus
  • Spell checker for Word & LibreOffice

Uyghur

  • Used in corpus-based phonology research

Tuvan

  • Used for UniMorph shared task (2021)
  • Spell checker for Word & LibreOffice

Sakha

  • Used in Revita (CALL)
  • Used for UniMorph shared task (2021)

Kyrgyz

  • Spell checker for Word & LibreOffice
  • Used for annotating published corpora

Developing a morphological transducer

lexd + twol + spellrelax method

  • lexd: specifies morphotactics and lexicon
  • twol: handles morphophonology
  • spellrelax: handles (~consistent) spelling variations

Apertium framework

  • bootstraps a vanilla template
  • includes compilation scripts
  • uses widely supported HFST binaries
  • lots of infrastructure built on it

Getting started

Concluding thoughts

  • Symbolic tools
    • are robust tools and shouldn't be ignored because of ML hype
    • (are one way to) enable meaningful langauge technology
  • YOU can develop symbolic tools

Thank you!

slides available at
https://jonorthwash.github.io/2026-LangTech-ClingDing/presentation.html

Case study: Turkic morphological transducers

Free/Open Source (versus proprietary)

Free/Open Source

Free = as in speech (and beer)
Open = made available (and insides showing)

What this means (ideally / in the case of Apertium):

  • Available for anyone to use, modify, or build on
  • Robust community, support from developers
  • Able to easily compare to other systems
  • Abandoned projects can be resumed or repurposed
  • Standards / tools for wider integration/use (e.g., Apertium: website platform, integration into OmegaT & Wikipedia Content Translation tool)

→ i.e., maximally accessible to language communities

Turkic: Language vitality

Turkic languages by number of speakers

>200 million speakers total

Language vitality: Turkic

Turkic languages by vitality

~40 languages, only 11 "normal" (Ethnologue 2023, other sources)

(estimates probably overly optimistic)

The full list

language№vitality
Turkish 88M normal
Uzbek 33M normal
Azərbaycani 24M normal
Kazakh 17M normal
Uyghur 12M normal
Turkmen 11M normal
Kyrgyz 7.3M normal
Tatar 5.5M normal
Iraqi Türkman 4.5M normal*
Bashqort 2M vulnerable
Chuvash 1.1M vulnerable
Qashqayi 1M normal
Khorasani 650K vulnerable
Qaraqalpaq 650K normal
Crimean Tatar 600K severely endangered
Sakha 500K vulnerable
Qumuq 450K vulnerable
Qarachay‑Balqar 350K vulnerable
Tuvan 300K vulnerable
Urum 200K definitely endangered
language№vitality
Gagauz 150K critically endangered
Siberian Tatar 100K definitely endangered
Noghay 100K definitely endangered
Dobrujan Tatar 70K severely endangered
Salar 70K vulnerable
Southern Altay 60K severely endangered
Northern Altay critically endangered
Khakas 50K definitely endangered
Khalaj 20K vulnerable
Äynu 6K critically endangered
Western Yugur 5K severely endangered
Shor 3K severely endangered
Dolgan 1K definitely endangered
Dukhan <500 critically endangered
Krymchak 200 critically endangered
Ili Turki 100 severely endangered
Tofa 100 critically endangered
Karaim <80 critically endangered
Chulym <50 critically endangered
Fu‑yü Gyrgys <10 critically endangered
total211M
 

(Ethnologue 2023 and other sources)

Existing Turkic Language technology

Mobile keyboards

(Google, Microsoft, Apple)

Machine translation

(Yandex, Google)

Text-to-speech support

(Android, Siri)

Baseline localisation data

(CLDR)
 

Existing Turkic Language technology

ISO 639 language codes

(baseline for any hope of language technology)

Currently missing: Iraqi Türkman, Fu-yü Gyrgys, Dukhan

  • Most of those are corporate efforts
  • What about non-corporate efforts?
    • academic
    • community-based
    • quite a few cross-over

Open/community-based efforts to develop Turkic language technology

TurkicInterLingua (TIL)

  • Machine Translation
  • Other NLP tasks

Apertium

  • Machine Translation
  • Morphological transducers

(case study, viz. next slides)

Universal Dependencies

Linguistically annotated text corpora

Kazakh, Kyrgyz, Tatar, Turkish (×4), Uyghur

Common Voice

Audio corpora (for speech recognition)

Uyghur (513h), Bashqort (279h), Uzbek (266h), Turkish (135h), Kyrgyz (49h), Tatar (34h), Chuvash (29h), Sakha (24h), Turkmen (7.8h), Kazakh (3.8h), Azərbaycani (1.5h)

not ready: Qaraqalpaq (28%), Tuvan (20%)

Tesseract

Optical Character Recognition (OCR)

Azərbaycani (×2), Kazakh, Kyrgyz, Tatar, Turkish, Uyghur, Uzbek (×2)

Practicality of community priorities

can be developed privileging local agency releasable as/in stand-alone product / sufficient? can be developed without large corpus integrable into popular platforms popular tools can be updated/corrected
search engine support ✔ ✘ / – ✘ ✘ ✘
voice assistants ✔ ✔ / ✘ ✘ ✘ ✘
localisation ✔ ✔ / ✘ ✔ ✘ ✘
spelling/grammar checking ✔ ✔ / ✘ ✔ ✔/✘ ✘
simple keyboards ✔ ✔ / ~✔ ✔ ✔/✘ ✘
predictive text keyboards ✔ ✔ / ~✔ ~✔ ✘ ✘
screen readers ✔ ✔ / ✔ ✔ ✘ ✘
MT ✔ ✔ / ✔ ✔ ✘ ✘
CALL ✔ ✔ / ✔ ✔ ✘ ✘

Practicality of community priorities

Privileging local agency

  • Everything for marginalised languages can be developed privileging local agency
  • Almost nothing actually is

Integration

  • Many tools are not sufficient as stand-alone products
  • It's almost impossible to integrate most tools into popular platforms
  • It's entirely impossible to update/correct popular tools

Privileging local agency

The specific challenges

  • Localisation: much software is not localisable independently of the producer of the software
  • Closed interfaces: language-related programming interfaces are often closed to developers
  • Closed resources: language resources are closed and not accessible to the public
  • Disregarded standards: language-related international standards are not respected or fully implemented
(Moshagen & Trosterud 2019)

Stepping back

Shuts out communities and those supporting them

… also just bad design (but intentionally so!)

Why???

Corporate profit models often demand this

Privileging local agency

What can be done?

  • Demand from your corporations (per Moshagen & Trosterud 2019)
    • Open localisation
    • Open interfaces
    • Open resources
    • Accessible standards
    Caveat: some communities have data sovereignty requirements (Mahelona 2020)
  • Support community priorities, privilege local agency
    (and ask your corporations to as well)
    Caveat: engagement with a community can take forms that do not privilege local agency (Long 2007, Romero 2016)
  • Focus on technology that can be meaningful without full integration
  • Close the loop: make the results of your work available to communities

Why?

  • Bring value to marginalised communities
  • Help prevent "digital language death"
  • Be an ethical language technologist