Regarding your example with Auschwitz: how exactly did it learn that Auschwitz belongs to concentration_camp type? Is it because your example sentence happened to say exactly that or is that just a coincidence?
Hi, great video! You've mentioned "... if you don't have training data". I am assuming that you mean that annotated data is not required, and instead the model relies on unsupervised approach. If this is correct, than for specialized texts it must rely on embedding training? Thanks!
Really cool! Can you make a video on how to further train the LatinCy model? I have a ton of additions to the lemma fixer custom component and I've noticed a few recurring patterns I want to fix generally
I can but it may be better to retrain from scratch. In these instances you can experience catastrophic forgetting. If you want to train from scratch, you could modify the original training data or add to it with your own. That said, if you simply need to adjust a component, that is entirely different. Would you mind explaining a bit more about what you want to do?
Let me start by saying I'm new to this! I put together a Latin corpus of texts and I'm counting lemma frequencies. But I noticed that some verb forms are consistently off, like almost all pluperfect forms (like counting uiderat as the lemma instead of uideo). Instead of having to add a correction to the component for each verb, I wanted to see if there was a way to train the model to make it better at recognizing the lemma of verbs in pluperfect forms. Thanks for responding!@@python-programming
@@JoseSanchez-xz5wt ahh I see! In that case, I would reach out to Patrick directly: twitter.com/diyclassics?lang=en --- he's on Twitter as diyclassics and if you look him up on Google you can find his email as well. I don't want to put it here and have him get spam messages.
No problem! Great question. I have not seen an example of this. Sorry! There are already a few examples of few-shot spaCy libraries. Concicse concepts is one.
Great video! What would you do to extract hard skills and soft skills from a resume and job description? I am thinking entity rulers from spacy and match it but I was wondering what you were thinking. Thanks!
Thanks! If you have a controlled vocab for these things then maybe a rules pipeline would work, but an ML model would likely be better since it wouls find things that are not in your list. You could also have a combination of both.
Hi thanks for the informative video! Let's say, like in your book, you had a list of concentration camps that you wanted to feed to the model to improve its accuracy. How would do that? Or would you not do it and just use a more conventional spaCy pipeline?
Thanks so much! Like most ML things, the best thing to do is try it out. Change the lagels to those exact label names and run it over a text. If you want to extract label names with spaCy, though I created bio spaCy that does precisely this.
@@python-programmingThank you so much, can you please share link to bio spacy? Additionally thank you so much for such amazing videos, they are really helpful!
@@ifrasaifi1124 That would be a separate component that does not yet exist. You want to look into entity linking and connecting the plant to a wiki_id and then connect that to a database of medicinal effects.
Great video. The only problem is that gliner is not easy to implement in production such as in a remote server or a huggingface endpoint. Has anyone able to make this work?
Thanks! Not sure if you were the one who left this comment on GitHub, but I just responded to an issue there. I'm curious if a good solution would be to upload a spaCy pipelien with gliner-spacy to HuggingFace. This would make it a standard spaCy token classification pipeline and then allow the HuggingFace endpoint to work. I haven't tested this, so I'd be curious to learn if it works! You will likely need to drop the gliner-spacy component into the repository.
It will, but the latency may be an issue. I'm not sure of anything that can do real-time NER the way in which you can get transcriptions in near real-time.
im starting to use Space for entity extraction from the content of my competitors on the serp for a keyword (I work on SEO) but the entities that extracts are very very weird, leaving behind some more important (I use it in Spanish). Gliner might help?
I had better results on GLiNER then on OpenAI 3.5 on zero-shot. A lot of False Positive. But at least we have what to filter later, and good is that it works very fast on low CPU needs. Still waiting for few-shot learning example, sure it will help a lot. Anyone tested domain-knowledge way of doing staff?
is there anything like this, but for text classification? e.g.: I have a list of labels (topics) and a list of texts. And it has to tell me what topics are mentioned in which text