You can add location information to your Tweets, such as your city or precise location, from the web and via third-party applications. You always have the option to delete your Tweet location history. Learn more
One of my worries about current trends in AI research is the reinforcement of the long-term cultural & linguistic dominance of the Anglosphere. It's not just that English is the exclusive language of research (which seems inevitable). NLP research and models focus on English.
In particular, the lack of pre-trained models for other languages than English is cutting off these languages and cultures from the next generation of AI applications. Languages like Mandarin will probably be fine, but those with small numbers of speakers won't be.
Super interesting! I also hear from researchers who study emotional decisionmaking that the dubious "emotions-on-your-face" cues researchers are trying to teach algorithms to recognize are very culture-dependent.
I agree! And it’s not just the active “emotions on the face” it’s also that words that describe emotions have a different meaning to every single person. Reductionistic oversimplification of behaviours (linguistically & actual) is the opposite to actual lived human behaviours.
in research i agree but i guess the knowledge gained from training wirg the english language allows faster development of applications for other languages as well, or am i wrong?
Not really: English is a bit of an outlier in terms of linguistic structure in that it has impoverished morphology. We don't even have a proper future tense.
I wonder if AI models could help save dying languages?
Unfortunately, no. The current AI methods require translations of same content in both languages to train the Machine Translation model. Most dying languages don't even have a script. How do we get millions of words & sentences in the dying languages?
Pre-trained models in other languages than English are indeed lacking; at Hugging Face we're rooting for multi-lingual and that's why we currently have 11 multi-lingual models out of 33:https://huggingface.co/transformers/multilingual.html …
I have friends researching with pt-br, and the multi-lingual BERT yields worse results than using the ULMFiT or ELMo fully trained only on a pt-br Wikipedia dump. Until now, we can't see the benefits of using BERT. We would need an expensive machine to pre-train it.
I worked on NLP/NLU in Persian language, the lack of resources and datasets (even unlabeled) makes it so hard that we spent over 60% of our time creating tools to mine / gather data. Nowadays I see Mozilla is gathering some open datasets which might make things better, but /
Without a commercial interest, it's hopeless.
Twitter may be over capacity or experiencing a momentary hiccup. Try again or visit Twitter Status for more information.
