grobid VS Deep Java Library (DJL)

Interesting concept. Grobid tries to do the same https://github.com/kermitt2/grobid

I would suggest using grobid - a pipeline for extracting scientific PDFs into a common XML format which can be easily parsed. Grobid has quite a nice mature REST API that I've used in some of my own projects. It parses references and matches them to their DOI using the CrossRef API with a reported 95% F1 score. This should make your job pretty simple as far as I can tell - all you'd need to do is run your papers through grobid and then build a citation graph by comparing document DOIs.

Just use science-parse or GROBID. They have been designed for that exact reason.

- I ended up using Grobid, which converts the PDF to a very detailed XML format. The format is not a word processing format though, but a format specifically for representing scientific documents. I don't know, if it would, for example, contain tags about bold or italicized text. The tool is working really well, but since you probably cannot use the output XML format directly, it will need some postprocessing, which would be relatively simple with XML parsing libraries.

A few years ago I evaluated a few open source tools. In the end focused on GROBID. As usual, it depends on the type of document whether it works well for your use-case. There is some focus on it being "fast" (if that is a concern).

It seems to have been picked up by Eclipse and there is also Oracle Labs' Tribuo and Deep Java Library. All seem active, but I don't know much about any of them. I agree it's probably best to follow the community and use a more popular tool like PyTorch.

Although it may be a bit more work, you can do both machine learning and AI in Java. If you are doing deep learning, you can use DeepJavaLibrary (I do work on this one at Amazon). If you are looking for other ML algorithms, I have seen Smile, Tribuo, or some around Spark.

Image preprocessing I know less about, but tokenization is something I've dealt with a bunch. There are a few options, either push the tokenizer into the ONNX model and use MS's ONNX Runtime extensions (we've used this when working with sentencepiece tokenizers), port the tokenizer entirely to Java (we did this for BERT), or use a sentencepiece or HF tokenizers wrapper directly (e.g. Amazon's DJL did this - HF, sentencepiece).

https://djl.ai/ seems very promising. I've played around with it quite a bit, not in real production though. It's a very well documented (https://d2l.djl.ai/) and active project, with Amazon working on it.

Here is link number 1 - Previous text "DJL"

To give a contrasting perspective, I think the Java ecosystem is much better suited for many data science tasks, and has a growing and well-maintained set of libraries for general purpose machine learning. I won't list them all, but TF-Java, DJL et al. have implementations of many modern architectures and there are a number of excellent libraries (CoreNLP, Lucene et al.) for working with text.

I think DJL also use use it for their tutorials - https://docs.djl.ai/jupyter/tutorial/01_create_your_first_network.html.

AWS Deep Learning more deep learning.