Since the seminal work of Mikolov et al. (2013), word vectors of log-bilinear SVMs have found their way into many NLP applications as an unsupervised measure of word relatedness.
Due to the rapid pace of research and the publish-or-perish mantra of academic publishing, word vector experiments contain undisclosed parameters, which make them difficult to reproduce.
In our work, we introduce the experiments and their parameters, compare the published experimental results with our own, and suggest default parameter settings and ways to make previous and future experiments easier to reproduce.
We show that the lack of variable control can cause up to 24% difference in accuracy on the word analogy tasks.
Návaznosti
MUNI/A/1076/2019, interní kód MU
Název: Zapojení studentů Fakulty informatiky do mezinárodní vědecké komunity 20 (Akronym: SKOMU)
Investor: Masarykova univerzita, Zapojení studentů Fakulty informatiky do mezinárodní vědecké komunity 20, DO R. 2020_Kategorie A - Specifický výzkum - Studentské výzkumné projekty
MUNI/A/1411/2019, interní kód MU
Název: Aplikovaný výzkum: softwarové architektury kritických infrastruktur, bezpečnost počítačových systémů, zpracování přirozeného jazyka a jazykové inženýrství, vizualizaci velkých dat a rozšířená realita.
Investor: Masarykova univerzita, Aplikovaný výzkum: softwarové architektury kritických infrastruktur, bezpečnost počítačových systémů, zpracování přirozeného jazyka a jazykové inženýrství, vizualizaci velkých dat a rozšířená realita., DO R. 2020_Kategorie A - Specifický výzkum - Studentské výzkumné projekty
NOVOTNÝ, Vít. The Art of Reproducible Machine Learning: A Survey of Methodology in Word Vector Experiments. In Aleš Horák, Pavel Rychlý, Adam Rambousek. Proceedings of the Fourteenth Workshop on Recent Advances in Slavonic Natural Language Processing, RASLAN 2020. Brno: Tribun EU, 2020, s. 55-64. ISBN 978-80-263-1600-8.
@inproceedings{1699699, author = {Novotný, Vít}, address = {Brno}, booktitle = {Proceedings of the Fourteenth Workshop on Recent Advances in Slavonic Natural Language Processing, RASLAN 2020}, editor = {Aleš Horák, Pavel Rychlý, Adam Rambousek}, keywords = {Machine learning; word vectors; word2vec; fastText; word analogy; reproducibility}, howpublished = {tištěná verze "print"}, language = {eng}, location = {Brno}, isbn = {978-80-263-1600-8}, pages = {55-64}, publisher = {Tribun EU}, title = {The Art of Reproducible Machine Learning: A Survey of Methodology in Word Vector Experiments}, url = {https://nlp.fi.muni.cz/raslan/raslan20.pdf#page=63}, year = {2020} }
TY - JOUR ID - 1699699 AU - Novotný, Vít PY - 2020 TI - The Art of Reproducible Machine Learning: A Survey of Methodology in Word Vector Experiments PB - Tribun EU CY - Brno SN - 9788026316008 KW - Machine learning KW - word vectors KW - word2vec KW - fastText KW - word analogy KW - reproducibility UR - https://nlp.fi.muni.cz/raslan/raslan20.pdf#page=63 N2 -
Since the seminal work of Mikolov et al. (2013), word vectors of log-bilinear SVMs have found their way into many NLP applications as an unsupervised measure of word relatedness.
Due to the rapid pace of research and the publish-or-perish mantra of academic publishing, word vector experiments contain undisclosed parameters, which make them difficult to reproduce.
In our work, we introduce the experiments and their parameters, compare the published experimental results with our own, and suggest default parameter settings and ways to make previous and future experiments easier to reproduce.
We show that the lack of variable control can cause up to 24% difference in accuracy on the word analogy tasks.
ER -
NOVOTNÝ, Vít. The Art of Reproducible Machine Learning: A Survey of Methodology in Word Vector Experiments. In Aleš Horák, Pavel Rychlý, Adam Rambousek. \textit{Proceedings of the Fourteenth Workshop on Recent Advances in Slavonic Natural Language Processing, RASLAN 2020}. Brno: Tribun EU, 2020, s.~55-64. ISBN~978-80-263-1600-8.