Towards the First Named Entity Recognition of Inuktitut for an Improved Machine Translation

Ngoc Tan Le, Soumia Kasdi, Fatiha SadatACL 2023

TL;DR

This paper introduces the first Named Entity Recognition (NER) system for Inuktitut, a low-resource Indigenous language. By leveraging transfer learning from English and building a novel annotated corpus, the authors demonstrate improved NER performance and a subsequent boost in Inuktitut-English Neural Machine Translation (NMT).

What Problem Does This Paper Address?

Named Entity Recognition (NER) is a fundamental task in Natural Language Processing (NLP), crucial for applications like machine translation and information retrieval. However, applying NER to low-resource languages, especially Indigenous languages like Inuktitut, is severely hampered by the lack of linguistic resources and annotated data. Inuktitut, spoken by approximately 39,770 people, faces the threat of extinction, making its revitalization and the development of NLP tools for it a critical endeavor. This paper aims to bridge this gap by developing an NER system for Inuktitut, thereby contributing to its preservation and enabling better NLP applications.

Methods

The authors propose two main approaches for Inuktitut NER, both relying on transfer learning from English, a high-resource language:

  • Rule-Based Approach: This method combines word alignment information with a morphological analyzer. The process involves extracting named entities from an English corpus, performing morphological analysis on Inuktitut text, identifying nominal groups, and then filtering these groups using word alignment to identify named entities. A bilingual dictionary and a knowledge base are also built to aid this process.
  • Bilingual Word Embedding-Based Approach: This approach utilizes unsupervised transfer learning based on bilingual word embeddings. It involves building an English-Inuktitut dictionary, recognizing English named entities, training monolingual word embeddings for both languages, and then projecting these embeddings into a shared space. The similarity between word vectors is used to identify named entities.

For the downstream task of improving Machine Translation (MT), they built an English-Inuktitut NMT framework. This framework leverages pre-trained word embeddings for initialization and incorporates source-target alignment information as an additional feature. Morphological complexity in Inuktitut is handled through morpheme segmentation.

Datasets

  • Nunavut Hansard Inuktitut-English Parallel Corpus 3.0: This corpus, containing proceedings of debates, was used for training and evaluating the proposed models. It comprises approximately 8,068,977 words in Inuktitut and 17,330,271 words in English, forming about 1.3 million sentence pairs.
  • UQAILAUT Project Database: This database was used to build a bilingual dictionary of 1,560 words, including root word meanings and suffix meanings. Microsoft Bing translator was used to translate frequent English words into Inuktitut.
  • Small Annotated Dataset for Inuktitut Named Entities: For evaluating the rule-based method, a small dataset was manually annotated, containing 4 types of named entities: LOC (45 entities), ORG (38 entities), PER (111 entities), and MISC (11 entities).
  • Bilingual Evaluation Dictionary: A dictionary of 30 word pairs was constructed to evaluate the translation performance in the common word embedding space.

Key Insights and Ideas

  • Transfer Learning is Viable for Low-Resource Indigenous Languages: The paper demonstrates that transfer learning from a high-resource language like English can be effectively applied to a low-resource language like Inuktitut for NER, even with significant linguistic differences.
  • Morphological Complexity is a Major Hurdle: Inuktitut's polysynthetic nature, where words can be formed by combining multiple morphemes, presents a significant challenge for word alignment and entity translation, often leading to misalignments and errors.
  • Rule-Based Methods Show Promise Despite Morphological Challenges: The rule-based approach, particularly for PER (person) entities, showed higher performance, suggesting that explicit linguistic rules can sometimes overcome the complexities that word embeddings might struggle with.
  • Word Embeddings Can Be Misled by Polysynthesis: The word-to-word translation approach using word embeddings sometimes distorts the translation of named entities, especially in polysynthetic languages where a single word can represent an entire sentence.
  • NMT Improvement through NER Integration: Incorporating NER capabilities into an NMT system can lead to tangible improvements in translation quality, as evidenced by the gains in BLEU, ChrF++, and TER scores.

Key Results

  • Word Alignment: The proposed Eflomal tool (ours) achieved a higher alignment error rate (AER) of 0.474 compared to other tools, indicating better performance in aligning words.
  • Rule-Based NER: The rule-based NER model achieved good performance across entity types, with Precision, Recall, and F1 scores ranging from 0.20 to 0.95, and notably high scores for PER (0.78 F1) and LOC (0.73 F1) entities.
  • Bilingual Word Embedding Translation: The word-to-word translation precision using bilingual word embeddings reached its best at 0.433 for a neighborhood of k=10.
  • NMT Performance: Model 1 (Transformer-based with word alignment) outperformed the baseline and Model 2 (Transformer-based with bilingual word embeddings) in terms of BLEU, ChrF++, and TER scores for English-Inuktitut translation. Specifically, Model 1 achieved a BLEU score of 32.84, compared to the baseline's 31.31 and Model 2's 31.70.

Limitations and Future Work

  • Limitations:
    • The rule-based approach's performance is affected by the morphological complexity of Inuktitut, which can mislead word alignment tools.
  • The word embedding-based approach can struggle with the polysynthetic nature of Inuktitut, leading to distorted translations of named entities.
  • The evaluation of the rule-based method used a small, manually annotated dataset.
  • Future Work:
    • The authors suggest exploring a hybridization of the rule-based and word embedding methods to leverage the strengths of both.
  • They aim to integrate knowledge bases related to toponymy and Indigenous knowledge into training word embeddings to improve system performance.
  • Future research will focus on emphasizing the differentiation between native Inuktitut named entities and borrowed ones.
  • Collaboration with an Indigenous community in Nunavut is planned to further develop NLP tools.

Summary updated

BibTeX
@inproceedings{le2023towards,
  author = {Ngoc Tan Le and Soumia Kasdi and Fatiha Sadat},
  title = {{Towards the First Named Entity Recognition of Inuktitut for an Improved Machine Translation}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2023},
}
Figure