I am a researcher in Natural Language Processing and I am bored at home. Anybody need help for their project?

Hey @zakke, sure. At my dayjob I work on both text analysis and NLG. My latest published paper is about syntax and social variation: here is the link if you are interested https://www.aclweb.org/anthology/P19-1246/.

I used to work with dyNet (http://dynet.io/) but since I joined my current company I started using Tensorflow directly and build what I need on top of it: the new Tensorflow (i.e. version > 2) changes everything and it really is a joy to use, compared to the hell it was.

For learning more: I think that the educational material that the spaCy people produce is great for developers getting started. If you already went through their tutorial, here are some resources that you migh find useful:

• follow @rctatman on twitter: she has some nice videos and recently started a 'NLP for developers' series
• Natural Language Processing with PyTorch by Delip Rao is a great book!
• https://www.youtube.com/watch?v=3Dt_yh1mf_U&list=PLQiyVNMpDLKnZYBTUOlSI9mi9wAErFtFm this is a traditional NLP course taught by Dan Jurafsky (author of a very popular NLP textbook) and Chris Manning: I can highly recommend this for approching somehow a classical trainining in NLP/computational linguistics
• Jaboc Eisenstein recently published a very nice textbook and he has a pre-print available on-line for free: https://github.com/jacobeisenstein/gt-nlp-class/blob/master/notes/eisenstein-nlp-notes.pdf

Hi there, thanks but I am not really looking for a new position. If there is any specific problem that you want to discuss here, I'd be happy to help if I can or point you to papers or resources that might help you. If you are looking for someone to hire, you should check https://nlppeople.com/, it's a job board for NLP/ML related positions.

Yeah, I'm indeed no expert in TTS and have no experience with any of these. Commercial models are harder to evaluate than academic research, but the nice thing is that they all usually let you play with their product for free a bit: It's usually a good idea to prepare a small test set and see what works best for you.

Hi @Corstian, I am sorry but I never worked on ASR or TTS and therefore I will not have anything really helpful to say. I've seen however at least one very good ASR model being released recently: https://paperswithcode.com/paper/listen-attend-and-spell. This model is an end-to-end network for ASR.

You could however try first with some SaaS solutions, to see if that's enough for your use-case:

• https://www.ibm.com/cloud/watson-speech-to-text
• https://aws.amazon.com/transcribe/
• https://cloud.google.com/speech-to-text

I've only played with the speech-to-text solution by Google and it was more than enough for my case. Maybe you can do some pre-processing to cut out some noise? And from whatI've seen dialects are starting to become a minor issue for ASR. Also, in the case of IBM Watson, you can also fine-tune the model to your data.

Hi there @flow34, thanks to you for your message and to all you guys here: you've managed to keep boredom away a bit :)

If I understand it correctly, you mostly want some statistics: I would give scattertext a try https://github.com/JasonKessler/scattertext and play with it a bit.

You might then want to look into systems for detecting multi-word expressions and unusually frequent phrases, which are the ones that your customers might be interested in.

If you need a quick way to plug-in sentiment analysis, my company recently opened a free API for that. Ping me in private if interested and I'll point you to it (I don't know what's the policy on IH in this case: is it ok to link stuff I am involved into?)

Hey @thatandyrose, that's a lot of stuff you put in here! :)

You are mostly talking about 'statistical NLP' here, which --- in simple terms --- is about learning from human-labelled examples how to do some task: the result of this learning is called a model. Machines can only process numbers and that's why you need to somehow map words to numbers and then you might want to find some patterns in those numbers that help you solving your task. An alternative approach is the 'rule-based NLP': spaCy's docs have a useful section on this https://spacy.io/usage/rule-based-matching. Once you have a (rule-based or statistical) model, evaluating it is a science of its own: but you usually define some metric that you care about (e.g. accuracy in the case of spam detection) and you have a human-labelled test set, i.e. you know already if a text is spam or not: you then run your two models on this dataset and the one that achieves the highest accuracy is the best one. I can highly recommend this book for getting started: http://themlbook.com/ (disclaimer: I was involved in the translation of the Italian edition).

Hi @Bsme, I see at least two ways to model this problem:

a) frame it as an information extraction problem. You could take a look at this page http://nlpprogress.com/english/information_extraction.html for learning about the possible approaches and what is the state-of-the-art
b) more simply, you could break it down into multiple, simpler problems, e.g.

• build a classifier for saying if the order was accepted/not-accepted
• train a Named Entitity Recognizer (NER) to extract things like order_date, etc. spaCy is usually a good solution for this: you might want to take a look at their tutorial on training a NER model on custom entities https://spacy.io/usage/training#ner

Hey there, question answering (QA) is a subfield of NLP which has started to gain a lot of attention recently. Take a look at this: https://rajpurkar.github.io/SQuAD-explorer/. It reports the best results on a question answering shared task with links to the relevant papers: it' s a good way to see what's the state-of-the-art on the task. Real life results tend however to be quite different from what you see on a research dataset.

The transformers library from huggingface made it possible to get started with QA super quickly thanks to their "pipeline" component: you should probably take a look at this: https://github.com/huggingface/transformers#quick-tour-of-pipelines.

Hi @nakurai, you could try to use word2vec (or something like BERT) to get the representations of the texts and then you would need:

• a clustering algorithm to group together texts that are somehow similar
• a way to automatically label the clusters

https://radimrehurek.com/gensim/ is usually a very good starting point for this: in particular, take a look at their tutorial on LDA. One thing though: this task with short texts is particularly challenging. But once you get started and have one baseline, then there are several things you can try to improve it.

Naming the clusters once you have them, e.g. having something like 'economy' is a though problem: maybe this paper will be helpful for you https://arxiv.org/abs/1612.05340. But you could also start labelling the clusters manually at the beginning.

Hey @fu_wire, spaCy is becoming the go-to solution for this. You probably want what's called a dependency parse of a text: https://spacy.io/usage/linguistic-features#dependency-parse

Hi @onraykimbo, if there is any specific problem that you can discuss here, I'd be happy to help if I can. NLTK is a good didactic tool but, yes, spaCy is probably what you need for building commercial products today. https://nlp.stanford.edu/ is also a good option, you might want to check that out.

Hi @Blake_Emigro, well no, there is a lot of research on voice as well, but that's not something I really know much about: the methods are much closer to digital signal processing and physics (for things like acoustic modelling). From what I've seen, since text and voice are everywhere, there are applications in every industry. If you are into those things, you should follow @solyarisoftware on twitter.

Among the freely available datasets, https://commoncrawl.org/ is the first one that comes to mind if you ask about size.

NLP is about text and therefore most of the time the average dataset is big enough to fit into memory.

With respect to your last question: not really, or not in the sense I think you mean it. Some times you do want to combine different datasets, but the text annotations don't match and normalizing them is not always straightforward.

With respect to accessing and using the different datasets, I am starting to see cool abstractions/libraries: my favourite is https://www.tensorflow.org/datasets.

Hi there, could you tell me something more? Do you already have a corpus of conversations that you want to model? I don't think you need sentiment analysis for this and intent detection usually refers to the task of understanding what a user wants to do with uttering a specific sentence, while here it seems you need more a higher level classification before you route the user to a building-habit or breaking-habit dialogue.