TL;DR
This paper summarizes the first workshop focused on applying modern language technology to the critical task of documenting and revitalizing endangered languages. It highlights the challenges of low-resource scenarios and presents prototypes for tools addressing speech, dictionaries, and educational resources, demonstrating progress in bridging the gap between advanced NLP and community needs.
What Problem Does This Paper Address?
The paper tackles the significant gap between the rapid advancements in natural language processing (NLP) and its limited application to the documentation and revitalization of the world's thousands of endangered languages. While NLP excels with high-resource languages like English or Mandarin, most languages lack the vast datasets (hundreds or thousands of hours of transcribed speech, large text corpora) required for traditional NLP models. This scarcity hinders efforts to create usable language technology tools for communities working to preserve their linguistic heritage, especially as fluent speakers age and the urgency to document their languages increases. The workshop aimed to bring together linguists, community members, and technologists to collaboratively develop practical, low-data solutions.
Methods
The workshop focused on four key areas, each with specific methodological approaches:
- Spoken Language Technology:
- First-Pass Approximate Transcription: Developed and improved "Allosaurus," an ASR system designed for low-resource settings. It uses a multilingual model trained on typologically diverse languages and leverages the PHOIBLE database to restrict output to phonemes present in the target language's inventory, significantly improving practical usability and reducing post-editing effort. An interface for customizing phonetic inventories was also created.
- Phone to Orthography Decoder: Explored "Higgins," an attempt to automate the conversion of phonetic transcriptions to native orthography. This involves using phonetic similarity metrics and adapted Viterbi algorithms, with initial results showing promise when applied to manual transcriptions.
- Text-to-Speech (TTS): Utilized unit selection TTS with the Festival system, demonstrating its feasibility even with limited data (e.g., a 3.8-hour corpus for SJQ Chatino), successfully handling tonal complexities.
- Text/Speech Forced Alignment: Adapted ReadalongStudio for cross-linguistic forced alignment, using pre-trained English models and PanPhon for sound mapping to align text and speech for SJQ Chatino, creating interactive children's books.
- Dictionary Management: Focused on converting unstructured dictionary data (from PDF/Word files) into structured, searchable formats. This involved using regular expressions for data extraction and creating a demonstrative website to showcase searchable dictionaries for Kwak'wala, Inuktitut, and Kidaw'ida, including translations, glosses, and morphological information.
- Teacher-in-the-Loop: Search Tools for Low-Resource Corpora: Developed a user-friendly search interface for language teachers. This system allows users to input example sentences and uses fuzzy string matching and word embeddings (fastText) to find semantically or orthographically similar sentences in an unannotated corpus, with a human-in-the-loop feedback mechanism to refine search results.
- Social Media: Explored the use of bots for language learning and analysis of social media text. This included developing an "Idiomata" bot for Seneca and Cayuga that tracks word usage in conversations and provides translations, incorporating fuzzy matching for error tolerance.
Datasets
The paper discusses the use of several datasets and resources:
- PHOIBLE database: Used to constrain phonetic output in the Allosaurus ASR system.
- PHOIBLE's Chatino inventory: Specifically used for demonstrating the Allosaurus customization feature.
- SJQ Chatino corpus: A 3.8-hour corpus of Eastern Chatino from the GORILLA language archive used for TTS development.
- English ARPABET inventory: Used in the ReadalongStudio forced alignment for SJQ Chatino.
- Microsoft Word and PDF files: Original formats for dictionaries of Kwak'wala, Inuktitut, and Kidaw'ida, which were converted to structured formats.
- Arapaho corpus: A 100,000-token corpus used for training word embeddings in the Teacher-in-the-Loop system.
- Bilingual dictionary (Seneca/Cayuga-English): Used as the basis for the Idiomata bot's translation and scoring functionalities.
- Irish texts: Used for social media text analysis and temporal language shift modeling.
No new benchmarks were explicitly introduced, but the work focused on adapting existing technologies to low-resource scenarios and creating functional prototypes for specific languages.
Key Insights and Ideas
- The critical role of phonetic inventories: Realizing that ASR systems need to be constrained by the actual phonemic inventory of a target language, rather than just a broad set of possible sounds, is crucial for practical transcription in low-resource settings.
- Bridging the gap between linguistics and NLP is essential: The workshop highlighted that successful language technology for documentation and revitalization requires close collaboration between community members, linguists, and computer scientists.
- Low-data NLP is achievable: The prototypes demonstrated that many advanced NLP tasks (ASR, TTS, forced alignment, search) can be adapted and made functional even with very limited data, often by leveraging cross-lingual transfer, existing linguistic databases, and clever algorithmic design.
- Interactive tools empower users: The development of user-friendly interfaces like the dictionary website, the Teacher-in-the-Loop search, and the Idiomata bot shows that technology can be made accessible and useful for language teachers and learners.
- The urgency of documentation: The paper implicitly underscores the critical need to develop these tools quickly, given the endangered status of many languages and the aging population of fluent speakers.
Key Results
- Improved Allosaurus transcription: Incorporating PHOIBLE inventories led to an 11-13% improvement in recognition accuracy and made human post-editing more practical.
- Functional TTS for SJQ Chatino: Excellent results were achieved with a 3.8-hour corpus, demonstrating the viability of unit selection TTS for tonal languages.
- Interactive dictionary website: A functional website was created for Kwak'wala, Inuktitut, and Kidaw'ida, allowing users to search for words and retrieve translations and other linguistic information.
- Teacher-in-the-Loop prototype: A working system was demonstrated for Arapaho, allowing teachers to find relevant sentences in an unannotated corpus using example queries.
- Idiomata bot: A functional chatbot was developed for Seneca and Cayuga, offering translation and conversation monitoring features with fuzzy matching.
Limitations and Future Work
- Orthographic decoding challenges: The "Higgins" decoder for phone-to-orthography conversion showed mixed results, indicating that further improvements are needed in both the ASR system and the decoder itself.
- Data scarcity for embeddings: Training effective word embeddings for low-resource languages remains a challenge due to the limited availability of large corpora. Future work aims to explore transfer learning and other advanced techniques.
- Cross-linguistic alignment limitations: The forced alignment for SJQ Chatino revealed that syllable structure, particularly initial nasal-plosive onsets, can pose challenges for models trained solely on English. Future work will investigate incorporating multilingual IPA-output models.
- Generalizability of tools: While prototypes were developed for specific languages, further work is needed to ensure their broad applicability across a wider range of languages and linguistic features.
- Community engagement: The organizers noted a "slant towards technologists" in the workshop and suggested that future workshops should be located in places more conducive to participation from language community members.
- Future directions: The paper suggests expanding dictionary annotations, incorporating more languages, developing better visualization for linguistic data, and further refining social media bots and search interfaces. A tutorial on technology for endangered languages is also planned for COLING 2020.
Summary updated