flair VS NLP-progress

The code of the module is available there for easy access: https://github.com/flairNLP/flair

Had the same experience with the german model in spacy (but tbh, the quailty of my textdata was bad). A bert based approach with flair really improved my results. I think there is a spanish pretrained model also available

We have a list of entities in text files separated with a new line. We intend to train the flair model to detect these entities in text, but NER models require the entity to be labeled in a paragraph with BOI format.

Training most of the Named Entity Recognition (NER) models for example Flair usually needs to format data in BOI tagging) scheme as shown below where each sentence is separated by blank line

I had the same problem a couple years ago. I think Flair, form Zalando uses a different Corpus. However, it's not great and I am pretty sure they are infringing the license anyway...

For example, my first approach to the pet sentences would be to label all sentences within a respective text corpus containing according information for either yes or no. You would then convert this to a tertiary tag set, something like ["pet allowed", "pet not allowed", "irrelevant"]. You could then try out a model based on SentenceBert, other sentence-level embeddings/language models or 1D CNNs for this. flairNLP (https://github.com/flairNLP/flair) is a small, little framework which provides comfortable high-level access to different common language models which integrates perfectly with pyTorch.

For NER, if you don't need the full toolkit of spacy, I'd highly recommend checking out Flair. It will likely run faster than transformer-based models (like en_core_web_trf) and it tends to be one of the best performing approaches to NER.

You then train the whole thing on sequences where each position has a label that is begin/inside/outside and thus you can calculate cross-entropy loss. So all in all it is basically: https://github.com/flairNLP/flair, https://huggingface.co/transformers/model_doc/distilbert.html#tfdistilbertforsequenceclassification or any huggingface model "for sequence classificaiton" or but just char based instead of word based. The CRF layer (as included in flair) is optional but may be useful.

One thing is that NLP is a set of wildly different problems which share some aspects, but often use quite different techniques and assumptions about their datasets. So even if you would have NLP experience, if you'd need to start on a substantially different NLP task, you can't just apply what you know and succeed, you have to review "how things are done" for that problem domain. For a quick overview, sites like https://nlpprogress.com/ can be helpful to see what methods are used; and, perhaps even more importantly, how people are modeling the actual task.

A standard step in Chinese text processing is word segmentation, which deals with this problem.

NLP has a github project tracking NLP progress, https://github.com/sebastianruder/NLP-progress. I wanna know if there is one tracking computer vision progress.

It would also be cool to have some charts with parameter density and even overall effectiveness (a tl;dr version of SOTA-trackers, maybe?) if that doesn't prove too infeasible.

As I said It increase ambiguity and cognitive overheard, needlessly given that "it" exists. Moreover it also make it harder for artificial intelligence to understand human text https://github.com/sebastianruder/NLP-progress/blob/master/english/coreference_resolution.md

I could not find it on the internet (including on GitHub, Kaggle, Medium, or Reddit.) And, I know about NLP Progress and The Super Duper NLP Repo.

For someone who needs to be on top of the latest research - Twitter (distraction-prone, marketing-friendly, instantly-gratifying, quick), newsletters in ML + NLP (https://jack-clark.net/, ruder.io, offconvex.org, etc.) (distraction-free, generic, time-consuming), SOTA chasing (https://paperswithcode.com/, http://nlpprogress.com/) (distraction-free, generic + focused, code-friendly)