Meta: Summary of future work ideas / feature requests

[jowagner]

Please directly edit this list rather than posting new comments. Each entry should link to an issue with more details. We can try to keep this list in order of priority.

• Compare to xml-r: issue #68
• Add an evaluation task focused on semantics: issue #84
• Train a roberta model: issue #8
• Effect of random initialisation (using the same model type and settings as in the development phase; purpose is to find out whether the observed differences between settings are meaningful; differences that can be explained with model instability and that contradict expectations can then be disregarded): issue #85
• Find out how much MT output is in the web-crawled corpora: issue #90
• Effect of filtering (near) duplicates: issue #73
• Add POS tagging as evaluation task: issue #92
• Effect of switching to v8 of Paracrawl: issue #77
• Train a GPT-2 model: issue #88
• For each corpus, effect of removing it
• Explore the sequence of data sets in https://github.com/jbrry/Irish-BERT/issues/72#issuecomment-830132236
• For each pair of corpora, effect of removing them
• More on filter thresholds and/or corpus-specific filter settings, see https://github.com/jbrry/Irish-BERT/issues/72#issuecomment-830013500
• Would CoNLL-style pre-tokenisation have any effect on Irish as it does for English, e.g. for don't it adds a boundary before n? Related: Issue #50
• Effect of increasing the weight of clean corpora: issue #53
• Restrict vocabulary building to clean corpora: issue #54
• Partition the corpus, create a vocabulary for each partition and join the vocabularies: issue #33
• Use lower BERT layer in parser: issue #93
• Effect of adding Lauren's and Teresa's Twitter corpora, in particular on tasks involving social media content: issue #34
• Add a copy of the data with accents removed and/or other normalisations: issue #30
• Add synthetic Irish, e.g. MT output: issue #14
• Add Hiberno-English corpora and English side of parallel corpora to training data: issue #13
• Add mC4 dataset: issue #89

TODO:

• There are more experiment ideas in issue #72