3
$\begingroup$

I want to match users to job posts, hoping to write a neural network to pick up on keywords users are looking for. My initial thinking was recurrent neural networks (RNNs) since we're dealing with language-modeling. But keyword positions shouldn't be important (keyword invariance). Eg, both the following job titles should match positive:

  • "Seeking React, React Native, Node, Postgres developer."
  • "Senior web developer. We use Node and React. Ideally also familiar with PostgreSQL."

My guess is that RNNs won't work here, and I should use CNNs instead since invariance is its specialty (at least with images). Alternatively, any other better-suited models than RNNs or CNNs?

$\endgroup$
  • $\begingroup$ Well, according to the book "Deep Learning" of Goodfellow et al. CNN's are appropriate for topologic-structured data (in example, images), while RNN's are the better choice when it comes to label sequence-structured data (in example, text). Can you elaborate what kind of structure your input has? $\endgroup$ – Unhandled exception Mar 10 '17 at 14:21
4
$\begingroup$

If order truly isn't necessary, recurrent neural networks might be the wrong tool to use altogether. I would recommend representing the job postings as bags-of-words or document vectors and performing another neural network architecture or simple nearest-neighbor search to start with.

$\endgroup$
  • $\begingroup$ Only recurrent neural networks assume an ordering in the inputs. Feedforward neural networks are just as capable as nearest-neighbor to work with bag-of-words. $\endgroup$ – Pieter Mar 10 '17 at 9:17
  • $\begingroup$ Starting simple though is a good advice. $\endgroup$ – Pieter Mar 10 '17 at 9:18
  • $\begingroup$ Yes but he specifically asked about CNNs, which have certain spacial arrangement/invariance assumptions baked into them, which don't apply to arbitrary BOW vectors. And if he were to use a fully-connected network instead, then a single network layer will need to contain ~vocab_size^2 many weights, which is likely to be impractical large. $\endgroup$ – jon_simon Mar 10 '17 at 20:01

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.