Beck, Nils Peer (2021)
Transfer Learning for Conceptual Metaphor Generation.
Technische Universität Darmstadt
doi: 10.26083/tuprints-00019359
Bachelor Thesis, Primary publication, Publisher's Version
|
Text
thesis_final.pdf Copyright Information: CC BY 4.0 International - Creative Commons, Attribution. Download (1MB) | Preview |
Item Type: | Bachelor Thesis | ||||
---|---|---|---|---|---|
Type of entry: | Primary publication | ||||
Title: | Transfer Learning for Conceptual Metaphor Generation | ||||
Language: | English | ||||
Date: | 2021 | ||||
Place of Publication: | Darmstadt | ||||
Collation: | 43 Seiten | ||||
Date of oral examination: | 6 July 2021 | ||||
DOI: | 10.26083/tuprints-00019359 | ||||
Abstract: | Metaphor can be understood as a linguistic phenomenon in which concepts from one domain – for instance, Money – behave the way concepts from a different domain – for instance, Liquid – usually do. “She drained her bank account” constitutes an example of a metaphorical expression, according to Conceptual Metaphor Theory. In this thesis, we address the task of metaphor generation, i.e., paraphrasing an input text into a metaphorical output text. First, we build a novel data set of more than 300,000 metaphorical sentence pairs, anchored in MetaNet and FrameNet, two lexical databases. Then, we fine-tune T5, a large pre-trained English language model, on metaphor generation in two different set-ups: Free metaphor generation, where the language model is expected to implicitly understand how to transform the input text, and controlled metaphor generation, where the language model is explicitly told what kind of metaphor to generate. We compare our fine-tuned models to related work in the field of metaphor generation, reporting promising results. While no clear favorite emerges along all evaluation criteria, our models perform particularly well on unseen metaphors. We also show that our free generation model produces more fluent and semantically similar text, while our controlled model produces more metaphorical text, suggesting differing use cases for both models. We make our code, our data set, and our fine-tuned models publicly available for further research. |
||||
Alternative Abstract: |
|
||||
Status: | Publisher's Version | ||||
URN: | urn:nbn:de:tuda-tuprints-193599 | ||||
Classification DDC: | 000 Generalities, computers, information > 004 Computer science 400 Language > 400 Language, linguistics |
||||
Divisions: | 20 Department of Computer Science > Ubiquitous Knowledge Processing | ||||
Date Deposited: | 10 Sep 2021 12:21 | ||||
Last Modified: | 15 Feb 2023 10:57 | ||||
URI: | https://tuprints.ulb.tu-darmstadt.de/id/eprint/19359 | ||||
PPN: | 485587912 | ||||
Export: |
View Item |