Coping with Zero-Shot Translation and its Explainability

在过去的几年里，人工智能的崛起令人瞩目。数据驱动的人工智能技术是最近取得的许多成功及其对我们社会的影响的幕后推手。 在与机器学习密切相关的人工智能的一个子领域--自然语言处理领域，神经机器翻译(NMT)在使用大量训练数据的情况下，在各种研究评估活动中获得了最先进的翻译质量。然而，这样的平行语料库并不是广泛适用于所有语言对和领域。因此，在翻译质量、处理资源非常少的语言和语言变体、考虑综合语言、处理词汇量不足、使用多种情态、翻译多个句子等方面有很大的改进空间。 我的研究计划的长期目标将解决用于建立高效和透明的零精度翻译系统的算法方法的稀缺性，并将处理原创性和挑战性的问题，这些问题将在适用于北美土著语言的机器翻译领域产生新的算法和技术。 我们打算在无监督的资源环境和可解释的环境中为无监督机器翻译的重要挑战做出贡献，同时创建新的翻译模型，除了与这两种语言相关的单语资源和知识表示外，什么都不依赖。此外，我们打算开发新的算法来解释中间结果，并为自动翻译系统提出的解决方案提供推理，从而在保持高水平学习性能的同时拆开黑盒。 拟议的研究发现方案将导致机器翻译和语言资源开发领域的革命性进展，并将有助于开发适用于土著语言的创新技术工具。此外，这一研究项目将使我能够巩固和扩大我在行业和国际上的合作，并将在UQAM的三个计算机科学研究生项目中，为培养自然语言处理和人工智能领域的高素质人才创造一个良好的研究环境。 该研究计划将考虑所有形式的多样性，以导致更有影响力的研究，促进更大的创新，并导致更高的绩效。由于这一拟议方案将侧重于将北美土著语言作为多种综合语言，这将有助于吸引土著社区和来自不同种族或文化背景的人才，加强大学研究，改善土著人民的生活质量，并为土著社区和整个加拿大更强劲的经济增长做出贡献。

In the last couple of years, the rise of Artificial Intelligence has been remarkable. The data-driven Artificial Intelligence techniques are behind so many of the recent successes and the effects they have on our society. In the field of Natural Language Processing (NLP), a sub-field of Artificial Intelligence, which has close ties with Machine Learning, Neural Machine Translation (NMT) has achieved state-of-the-art translation quality in various research evaluations campaigns when using large amounts of training data. However, such parallel corpora are not widely available for all language pairs and domains. Thus, there is a lot of room for improvement regarding the quality of translation, processing very low-resourced languages and language variants, considering polysynthetic languages, handling the out-of vocabulary, using multiple modalities, translating multiple sentences, , etc. The long-term objective of my research program will deal with the scarcity of algorithm methods for building efficient and transparent zero-shot translation systems and will handle original and challenging problems that give rise to novel algorithms and techniques in the field of machine translation applied to North American indigenous languages. We intend to contribute to the important challenge of unsupervised machine translation in zero-shot resource settings and an explainable environment, while creating new translation models relying on nothing but monolingual resources and knowledge representations related to this pair of languages. Moreover, we intend to develop new algorithms that explain intermediate outcomes and provide reasoning for the proposed solutions of the automatic translation system, thus unpacking the black box while maintaining a high level of learning performance. The proposed research Discovery program will lead to revolutionary advances in the area of machine translation and language resources development and will contribute to the development of innovative technological tools applied to indigenous languages. Moreover, this research program will allow me to consolidate and expand my industrial and international collaborations and will create an excellent research environment for training Highly Qualified Personals in the area of Natural Language Processing and Artificial Intelligence, within three graduate programs in computer science at UQAM. This research program will consider all forms of diversity to lead to higher-impact research, fosters greater innovation, and results in enhanced performance. As this proposed program will focus on North American indigenous languages as polysynthetic languages, this will help attract talents in indigenous communities and from different ethnic or cultural backgrounds, to strengthen university research, to improve the quality of life for indigenous people and to contribute to stronger economic growth in indigenous communities, and in Canada as a whole.