Understanding the structure and meaning of Finnish texts: From corpus creation to deep language modelling
Kanerva, Jenna (2024-03-15)
Understanding the structure and meaning of Finnish texts: From corpus creation to deep language modelling
(15.03.2024)
Turun yliopisto
Julkaisun pysyvä osoite on:
https://urn.fi/URN:ISBN:978-951-29-9623-0
https://urn.fi/URN:ISBN:978-951-29-9623-0
Tiivistelmä
Natural Language Processing (NLP) is a cross-disciplinary field combining elements of computer science, artificial intelligence, and linguistics, with the objective of developing means for computational analysis, understanding or generation of human language. The primary aim of this thesis is to advance natural language processing in Finnish by providing more resources and investigating the most effective machine learning based practices for their use. The thesis focuses on NLP topics related to understanding the structure and meaning of written language, mainly concentrating on structural analysis (syntactic parsing) as well as exploring the semantic equivalence of statements that vary in their surface realization (paraphrase modelling). While the new resources presented in the thesis are developed for Finnish, most of the methodological contributions are language-agnostic, and the accompanying papers demonstrate the application and evaluation of these methods across multiple languages.
The first set of contributions of this thesis revolve around the development of a state-of-the-art Finnish dependency parsing pipeline. Firstly, the necessary Finnish training data was converted to the Universal Dependencies scheme, integrating Finnish into this important treebank collection and establishing the foundations for Finnish UD parsing. Secondly, a novel word lemmatization method based on deep neural networks is introduced and assessed across a diverse set of over 50 languages. And finally, the overall dependency parsing pipeline is evaluated on a large number of languages, securing top ranks in two competitive shared tasks focused on multilingual dependency parsing. The overall outcome of this line of research is a parsing pipeline reaching state-of-the-art accuracy in Finnish dependency parsing, the parsing numbers obtained with the latest pre-trained language models approaching (at least near) human-level performance.
The achievement of large language models in the area of dependency parsing— as well as in many other structured prediction tasks— brings up the hope of the large pre-trained language models genuinely comprehending language, rather than merely relying on simple surface cues. However, datasets designed to measure semantic comprehension in Finnish have been non-existent, or very scarce at the best. To address this limitation, and to reflect the general change of emphasis in the field towards task more semantic in nature, the second part of the thesis shifts its focus to language understanding through an exploration of paraphrase modelling. The second contribution of the thesis is the creation of a novel, large-scale, manually annotated corpus of Finnish paraphrases. A unique aspect of this corpus is that its examples have been manually extracted from two related text documents, with the objective of obtaining non-trivial paraphrase pairs valuable for training and evaluating various language understanding models on paraphrasing. We show that manual paraphrase extraction can yield a corpus featuring pairs that are both notably longer and less lexically overlapping than those produced through automated candidate selection, the current prevailing practice in paraphrase corpus construction. Another distinctive feature in the corpus is that the paraphrases are identified and distributed within their document context, allowing for richer modelling and novel tasks to be defined.
The first set of contributions of this thesis revolve around the development of a state-of-the-art Finnish dependency parsing pipeline. Firstly, the necessary Finnish training data was converted to the Universal Dependencies scheme, integrating Finnish into this important treebank collection and establishing the foundations for Finnish UD parsing. Secondly, a novel word lemmatization method based on deep neural networks is introduced and assessed across a diverse set of over 50 languages. And finally, the overall dependency parsing pipeline is evaluated on a large number of languages, securing top ranks in two competitive shared tasks focused on multilingual dependency parsing. The overall outcome of this line of research is a parsing pipeline reaching state-of-the-art accuracy in Finnish dependency parsing, the parsing numbers obtained with the latest pre-trained language models approaching (at least near) human-level performance.
The achievement of large language models in the area of dependency parsing— as well as in many other structured prediction tasks— brings up the hope of the large pre-trained language models genuinely comprehending language, rather than merely relying on simple surface cues. However, datasets designed to measure semantic comprehension in Finnish have been non-existent, or very scarce at the best. To address this limitation, and to reflect the general change of emphasis in the field towards task more semantic in nature, the second part of the thesis shifts its focus to language understanding through an exploration of paraphrase modelling. The second contribution of the thesis is the creation of a novel, large-scale, manually annotated corpus of Finnish paraphrases. A unique aspect of this corpus is that its examples have been manually extracted from two related text documents, with the objective of obtaining non-trivial paraphrase pairs valuable for training and evaluating various language understanding models on paraphrasing. We show that manual paraphrase extraction can yield a corpus featuring pairs that are both notably longer and less lexically overlapping than those produced through automated candidate selection, the current prevailing practice in paraphrase corpus construction. Another distinctive feature in the corpus is that the paraphrases are identified and distributed within their document context, allowing for richer modelling and novel tasks to be defined.
Kokoelmat
- Väitöskirjat [2881]