https://github.com/maaarcocr/papers-i-wrote
https://github.com/maaarcocr/papers-i-wrote
Last synced: 4 months ago
JSON representation
- Host: GitHub
- URL: https://github.com/maaarcocr/papers-i-wrote
- Owner: Maaarcocr
- Created: 2018-04-21T20:52:32.000Z (about 8 years ago)
- Default Branch: master
- Last Pushed: 2019-10-31T12:21:38.000Z (over 6 years ago)
- Last Synced: 2025-09-01T10:59:13.865Z (9 months ago)
- Size: 854 KB
- Stars: 0
- Watchers: 1
- Forks: 0
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
Awesome Lists containing this project
README
# papers-i-wrote
This is a collection of papers I wrote (or co-wrote)
## [Understanding language - one evaluation at a time](https://github.com/Maaarcocr/papers-i-wrote/blob/master/understanding-language-one.pdf)
In this survey I will firstly introduce the reader to the basic concepts that power the modern NLP research, describing concepts like neural networks, optimizations techniques and word embeddings. Then, I will describe the status of the art of evaluation in the NLP community, what problems exist and solutions. I will conclude exploring by future directions in evaluating NLP systems.
## [Hyper-parameter tuning and empirical evaluation in Natural Language Processing](https://github.com/Maaarcocr/papers-i-wrote/blob/master/hyper-parameter-tuning-and-evaluation.pdf)
As the field of Natural Language Processing (NLP) continues to mature, with more claimed state-of-the-art results, it has become increasingly difficult to discern between model architecture strength and irrelevant factors. We see our works being developed into a fully integrated framework, which would allow us to compare architectures on different data-sets with automatic hyper-parameter tuning. We introduce a systematic and reproducible method, which allows us to reduce bias. Our black-box optimisation not only leads to computational speedup, but also achieves stronger results than the word2vec default hyper-parameters; achieving a significant 15.9% improvement on a word similarity evaluation metric. We aim to usher in a new trend where future research has greater emphasis on careful empirical experimentation, leading to superior architectures.
## [”What does it refer to?” Casting Reference Resolution as Question Answering](https://github.com/Maaarcocr/papers-i-wrote/blob/master/what-does-it-refer-to.pdf)
Most approaches to coreference resolutionwhich are present in the current literature donot make extensive use of large pre-trainedlanguage models representations. We presenttwo new such approaches, by casting corefer-ence resolution as question answering and byusing a pre-trained BERT model fine-tuned onthe GAP dataset. This dataset was created inresponse to presence of gender bias in previousdatasets, thus it is a gender balanced corefer-ence resolution dataset. We achieve compet-itive results in this dataset under 2 differentvariations of the task it proposes. Our best re-sult is 84.7% accuracy.