Extracting useful information from Job description

[ryanwesslen]

hi @Aditya_varma_10!

Thanks for your question and welcome to the Prodigy community

@Aditya_varma_10

In order to classify sentences should I write seperate script to divide the whole document into sentences and create the dataset where each sentence in the document is single input for text classification to annotate using prodigy .

Typically, yes. The docs cover this:

If your documents are longer than a few hundred words each, we recommend applying the annotations to smaller sections of the document. Often paragraphs work well. Breaking large documents up into chunks lets the annotator focus on smaller pieces of text at a time, which helps them move through the data more consistently. It also gives you finer-grained labels : you get to see which paragraphs were marked as indicating a label, which makes it much easier to review the decisions later.

It's important to know that manual recipes do not do sentence segmentation be default, teach and correct recipes will do sentence segmentation by default. You can also use split_sentences too.

@Aditya_varma_10

I tried training a spancat model on whole document and the spans extend from 1 word to 20 words per span. I tried training the model and observed that training is taking too long to complete even after gpu.

Is there a better way to speed up the training process?

What suggester function are you using? If you're using the default ngrams then yes, this can drastically slow down training, especially on very long text.

Also this post suggests ways to break down the problem:

spancat with really large spans? (Identify sections in text)

I found a dialogue on the forum that might be inspirational here. It's a different problem, but it highlights another two-step approach to rethinking spans. That said, reading your reply still makes me think that textcat might be the simplest way forward, albeit on paragraphs instead of sentences. While I like your idea of using NER to detect the start of a section, I wonder if you might be able to leverage that this always starts on a newline, which suggests a heuristic might be better than…

Also, try running data debug (e.g., after running data-to-spacy) to get stats about your spans lengths:

What is the maximum length the annotation of a span can be? · explosion/spaCy...

@Aditya_varma_10

Is it better to use spancat instead of above approach to annotate sentence as both spans and entities.

Sorry, I don't understand this comment. Are you asking whether it is better to use spancat instead of ner? If so, the answer is it "depends".

(Also, that post was originally from 2019 and was before spancat was introduced.)

There's no hard rule but the Prodigy docs give you some idea:

Named Entity Recognition	Span Categorization
spans are non-overlapping syntactic units like proper nouns (e.g. persons, organizations, products)	spans are potentially overlapping units like noun phrases or sentence fragments
model predicts single token-based tags like B-PERSON with one tag per token	model predicts scores and labels for suggested spans
takes advantage of clear token boundaries	less sensitive to exact token boundaries

A good rule of thumb is if the meaning of your spans changes if you were to rearrange the words, then use ner. Alternatively, if you have more flexibility that you can rearrange the words within a span and it can have generally the same meaning, then spancat may work better.

Also, if you haven't seen, we have a case study project where we compared ner vs. spancat performance.

@Aditya_varma_10

If the training is taking too long is there a way to resume the training from last checkpoint in case the training is stopped before completion.

Yes, see this post:

Is It Possible to Resume Training Via CLI in Spacy v3 (transformers)? ·...

FYI, for spaCy specific questions (e.g., training, compute, GPU), I recommend searching/using the spaCy GitHub discussions forum.