Logo Logo
Hilfe
Kontakt
Switch language to English
Multilingual representations and models for improved low-resource language processing
Multilingual representations and models for improved low-resource language processing
Word representations are the cornerstone of modern NLP. Representing words or characters using real-valued vectors as static representations that can capture the Semantics and encode the meaning has been popular among researchers. In more recent years, Pretrained Language Models using large amounts of data and creating contextualized representations achieved great performance in various tasks such as Semantic Role Labeling. These large pretrained language models are capable of storing and generalizing information and can be used as knowledge bases. Language models can produce multilingual representations while only using monolingual data during training. These multilingual representations can be beneficial in many tasks such as Machine Translation. Further, knowledge extraction models that only relied on information extracted from English resources, can now benefit from extra resources in other languages. Although these results were achieved for high-resource languages, there are thousands of languages that do not have large corpora. Moreover, for other tasks such as machine translation, if large monolingual data is not available, the models need parallel data, which is scarce for most languages. Further, many languages lack tokenization models, and splitting the text into meaningful segments such as words is not trivial. Although using subwords helps the models to have better coverage over unseen data and new words in the vocabulary, generalizing over low-resource languages with different alphabets and grammars is still a challenge. This thesis investigates methods to overcome these issues for low-resource languages. In the first publication, we explore the degree of multilinguality in multilingual pretrained language models. We demonstrate that these language models can produce high-quality word alignments without using parallel training data, which is not available for many languages. In the second paper, we extract word alignments for all available language pairs in the public bible corpus (PBC). Further, we created a tool for exploring these alignments which are especially helpful in studying low-resource languages. The third paper investigates word alignment in multiparallel corpora and exploits graph algorithms for extracting new alignment edges. In the fourth publication, we propose a new model to iteratively generate cross-lingual word embeddings and extract word alignments when only small parallel corpora are available. Lastly, the fifth paper finds that aggregation of different granularities of text can improve word alignment quality. We propose using subword sampling to produce such granularities.
Not available
Jalili Sabet, Masoud
2022
Englisch
Universitätsbibliothek der Ludwig-Maximilians-Universität München
Jalili Sabet, Masoud (2022): Multilingual representations and models for improved low-resource language processing. Dissertation, LMU München: Fakultät für Mathematik, Informatik und Statistik
[thumbnail of Jalili_Sabet_Masoud.pdf]
Vorschau
PDF
Jalili_Sabet_Masoud.pdf

3MB

Abstract

Word representations are the cornerstone of modern NLP. Representing words or characters using real-valued vectors as static representations that can capture the Semantics and encode the meaning has been popular among researchers. In more recent years, Pretrained Language Models using large amounts of data and creating contextualized representations achieved great performance in various tasks such as Semantic Role Labeling. These large pretrained language models are capable of storing and generalizing information and can be used as knowledge bases. Language models can produce multilingual representations while only using monolingual data during training. These multilingual representations can be beneficial in many tasks such as Machine Translation. Further, knowledge extraction models that only relied on information extracted from English resources, can now benefit from extra resources in other languages. Although these results were achieved for high-resource languages, there are thousands of languages that do not have large corpora. Moreover, for other tasks such as machine translation, if large monolingual data is not available, the models need parallel data, which is scarce for most languages. Further, many languages lack tokenization models, and splitting the text into meaningful segments such as words is not trivial. Although using subwords helps the models to have better coverage over unseen data and new words in the vocabulary, generalizing over low-resource languages with different alphabets and grammars is still a challenge. This thesis investigates methods to overcome these issues for low-resource languages. In the first publication, we explore the degree of multilinguality in multilingual pretrained language models. We demonstrate that these language models can produce high-quality word alignments without using parallel training data, which is not available for many languages. In the second paper, we extract word alignments for all available language pairs in the public bible corpus (PBC). Further, we created a tool for exploring these alignments which are especially helpful in studying low-resource languages. The third paper investigates word alignment in multiparallel corpora and exploits graph algorithms for extracting new alignment edges. In the fourth publication, we propose a new model to iteratively generate cross-lingual word embeddings and extract word alignments when only small parallel corpora are available. Lastly, the fifth paper finds that aggregation of different granularities of text can improve word alignment quality. We propose using subword sampling to produce such granularities.