Language technology tools to support low-resource languages: case study of Sakha


Sardana Ivanova

2022-05-16

Contents

  • Sakha language
  • A free/open-source morphological analyser and generator for Sakha
  • Sakha in Revita—computer-assisted language learning platform
  • Poetry generation
  • Template-based news generation
  • Work in progress: Turkic benchmark
  • Conclusions

Sakha

  • Turkic language
  • ~450K speakers
  • Spoken primarily in Siberia
  • Official status in Sakha Republic
  • Entire area prone to economic and cultural impacts of climate change
  • Agglutinating — words may be inflected using a series of affixes
  • Backness and rounding vowel harmony systems
  • Word order: Subject-Object-Verb
  • Lena Turkic:
    • dlg - Dolgan
    • sah - Sakha
  • Sayan Turkic:
    • dkh - Dukha
    • kim - Tofa
    • tuh - Tuha
    • tyv - Tuvan
  • Yeniseian Turkic:
    • cjs - Shor
    • clw - Chulym
    • fyg - Fu-yü Gyrgys
    • kjh - Khakas
    • tuh - Tuha
  • Altay Turkic:
    • alt - Altay, Teleut, Telengit
    • atv - Chalqan, Qumandy, Tuba
    • tlt - Bachat Teleut
  • Kypchak (NW) Turkic:
    • sty - Siberian Tatar

A free/open-source morphological analyser and generator for Sakha

Joint work with


Jonathan N. Washington¹

Francis M. Tyers²,³
  1. Swarthmore College, USA
  2. Indiana University, USA
  3. Высшая Школа Экономики, Москва

Morphological tranducers

  • Twofold function: morphological analysis and generation
    атын ↔ атnpx3sgacc/атынadj/атынpost
  • Implemented as finite state transducers (FST)
  • Compiled from hand-coded lexical, morphotactic, and morphophonological generalisations
  • Only one development cycle
  • many uses in language technology and ``downstream'' tasks:
    • can be repurposed as spell checkers
    • may be used in rule-based machine translation pipelines
  • Some current uses of this transducer
    • used in Revita—language learning platform for revitalization and support of endangered languages (Katinskaia et al., 2018)
    • used to generate data in recent shared task on morphological reinflection (Pimentel et al., 2021)

Existing Turkic transducers

Comparison of lexicon size

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, Turkmen, Iraqi Türkman, Noghay, Khakas, Altay, Ottoman)
(Washington et al., 2019/2021), github.com/apertium
Sakha: >37 500 stems

Implementation and Challenges

Two-level approach using HFST

orthographic form

дьиэлэрбиттэн

↕ (twol: phonological mappings)

morphological/lexical form

дьиэ>LAр>BI>ттAн

↕ (lexc: lexicon + morphotactics)

analysis: lemma, POS, grammatical tags

дьиэnplpx1sgabl

Implementation and Challenges

Vowel harmony with diphthongs and long vowels

Problem:

  • Sakha long vowels (II, AA) behave like short-vowel counterparts
  • But diphthongs (IA) behave like high vowels (I)
    (round after any round V, do not trigger rounding of low Vs)

Solution:

  • Each twol harmony rule (char-to-char mapping): sensitive to whether harmonising V is component of long V or diphthong or not
  • Many V harmony alternations required multiple rules to implement

Implementation and Challenges

Two-directional consonant assimilation

Problem:

  • forms like /tutn-bIt-A/ ‘use-past-3’ realised as [tutummuta]
    тутyн>BIт>A:тутуммута
  • /n/ triggers nasalisation of the following /b/
  • /b/ triggers labialisation of the preceding /n/

Solution:

  • Mutual influence not problematic in twol
  • rules are sensitive to underlying form (left side of :) of adjacent symbol, not surface form (right side of :)

Implementation and Challenges

Many alternations in a single stem

Problem:

  • forms like уһун ‘swim[imp]’ / устар ‘swim-pres’
  • Several different alternations involved:
    • с ~ Һ — intervocalic lenition
    • н ~ т — sonority restrictions
    • I ~ ∅ — consonant cluster restrictions

Solution:

  • ≥1 twol mapping for each alternation
  • each mapping sensitive to the others & to other parts of morphophonological context
  • y used for high vowel ~ ∅ alternation, as in previous work (Washington et al., 2019)

Implementation and Challenges

Novel grammatical understanding

Existing literature:

  • Sakha exhibits many non-finite verb forms
  • Some have finite uses
  • Generally categorised roughly as "participle" or "converb"

Our contribution:

  • Categorised each form carefully based on uses:
    verbal noun, verbal adjective, verbal adverb, infinitive
  • Implemented each use separately
  • Results in some syncretism (forms existing across multiple categories)
  • Concluded that there is not a strict participle/converb dichotomy
  • Documented in more detail in Washington et al. (2021)

Evaluation

Coverage

  • Naïve coverage: the number of forms in a corpus that receive an analysis, regardless of whether or not the analysis is correct (e.g., in context)
Corpus Tokens Coverage
Newspapers ~16M 91.04%
Wikipedia ~2.4M 91.30%
New Testament ~190K 94.53%

Over 90% coverage: robust morphology

Evaluation

Precision & Recall

  • created gold standard:
    • 1000 valid words of Sakha
    • randomly selected from Wikipedia corpus
    • manually annotated output of transducer
Results:
Corpus Precision Recall
Wikipedia 98.52% 75.42%

i.e., nearly every form returned by the transducer was deemed correct, but many correct analyses were not returned by the transducer (mostly due to low coverage)

Future work

  • Correct minor issues in implementation of some morphophonological alternations
    identified recently in data generation for a shared task (Pimentel et al., 2021)
  • Morphological/syntactic disambiguation
  • More language technology applications of transducer? (spell checkers, MT, etc.?)

Conclusion

  • Robust transducer
    • high coverage
    • high precision
    • moderate recall
  • Lots of room for improvement!
  • Ready for use in language technology applications, downstream tasks
  • This work has also contributed to documentation of Sakha grammar

Revita

Computer-Assisted Language Learning system

A project of the Language Learning Lab at the University of Helsinki

Revita

Computer-Assisted Language Learning system

  • User uploads arbitrary text
  • System generates exercises
  • User practices with exercises
  • System gives feedback

Revita modules

Conclusion

  • The functionality of the language learning system is under development. For languages with more speakers many more linguistic resources and tools are available than for Sakha.
  • For example, currently, the Sakha system has only noun–postposition government rules.

Poetry generation

Joint work with

  • Michele Boggia
  • Anna Kantasalo
  • Simo Linkola
  • Hannu Toivonen

Poetry

There Will Come Soft Rains

Sara Teasdale - 1884-1933

There will come soft rains and the smell of the ground,
And swallows circling with their shimmering sound;

And frogs in the pools singing at night,
And wild plum trees in tremulous white,

Robins will wear their feathery fire
Whistling their whims on a low fence-wire;

And not one will know of the war, not one
Will care at last when it is done.

Not one would mind, neither bird nor tree
If mankind perished utterly;

And Spring herself, when she woke at dawn,
Would scarcely know that we were gone.

From The Language of Spring, edited by Robert Atwan, published by Beacon Press, 2003.

Interaction pattern

Both models are fine-tuned mBART models

Data

  • English
    • Gutenberg project
  • Finnish
    • Gutenberg project
    • Wikisource
    • Modern poetry from Poesia publishing house
  • Swedish
    • Wikisource

Conclusion

  • Simple co-creative poetry generator system
  • The highly limited interaction on purpose
  • The system generates lines using fine-tuned mBART models
  • The next step is an evaluation of the system and measures with actual users

Template-based news generation

Part of the Embeddia project

Cross-Lingual Embeddings for Less-Represented Languages in European News Media

Template-based news generation

CPHI (Harmonized Index of Consumer Prices)

Template-based news generation

Template example

Template-based news generation

Result example

В Апреле 2021, в Финляндии, ежемесячный темп роста согласованного индекса потребительских цен для категории 'здоровье' был 2.5 пункта. Ежемесячный темп роста согласованного индекса потребительских цен для категории 'здоровье' был на 2.3 процентных пунктов больше, чем в среднем по ЕС. В Марте 2021, ежемесячный темп роста согласованного индекса потребительских цен для категории 'здоровье' был на 2.2 процентных пунктов больше, чем в среднем по ЕС. Ежемесячный темп роста согласованного индекса потребительских цен для категории 'здоровье' был 2.3 пункта. Финляндия имела 2 самый высокий ежемесячный темп роста согласованного индекса потребительских цен для категории 'здоровье' во всех наблюдаемых странах. В Апреле 2021, Финляндия имела 1 самый высокое значение во всех наблюдаемых странах.

Turkic benchmark

A project of Turkic Interlingua community

The goal is to create benchmarks for Extractive Summarization, SQuAD Question Answering, NLI, NER, and some other tasks for 23 Turkic languages

List of languages

  • Altai
  • Azerbaijani
  • Bashkir
  • Chuvash
  • Crimean Tatar
  • Gagauz
  • Iraqi Turkmen
  • Karachay-Balkar
  • Karakalpak
  • Kazakh
  • Khakas
  • Kirghiz
  • Kumyk
  • Sakha
  • Salar
  • Shor
  • Tatar
  • Turkish
  • Turkmen
  • Tuvinian
  • Urum
  • Uyghur
  • Uzbek

Conclusions

  • A free/open-source morphological analyser and generator for Sakha
  • Sakha in Revita—computer-assisted language learning platform
  • Poetry generation
  • Template-based news generation
  • Work in progress: Turkic benchmark

Махтал! / Maxtal! / Thank you!