syw2014 / deepnlp-models

A repositiory about the practical deep learning models in NLP tasks, all of those models will be implement with tensorflow and test with chinese industrial dataset.

DeepNLP-Models

Table of Contents

1. Introduction

2. NLP Tasks

   • Text Matching
   • Text Search
   • Keyword Search
   • Document Classification

3. Learning to Rank

   • DeepFM

Introduction

As we all know deep learning has shown super power in image processing, so a lot researchers try to study NLP tasks with deep learning, and a various of deep nlp models haven discussion, that's what I want to do to collect the deep nlp models which can get the state-of-art performance, so that we can use those models in industry.

NLP Tasks

This part will introduce useful deep learning models in nlp tasks, and we try to make it available in industry. And the models will contain basic tasks like classification, text similarity matching, named entity recognition, ctr and some higher models(seq2seq)

Text Matching

Text matching are wide used in infromation retrieval, question answering, conversational response ranking etc, so we choose some simple and easy models, and implement them in an enffiency way with tensorflow. As a start, I will implement a baseline with DSSM
and the introduction about this model in here.

Text Search

Here we try to complete this task with another approach called $Deep Search$. In the current version we havn't add complicated algorithms, we just use word embedding and Faiss, encoding query and answers to embedding, and search with vectors, faiss speed search related answers. This may also can be a simple way to do QA, but the result may not very good.Check here to find module.

• Add implementation ideas of this parts
• Collect more QA data
• Implement service for Query search.

Keyword Search

Keyword search also call string match, it's a classical problem, in our scenario we use keyword search for text detect that we want to find which text or document can not be showing to user.I have implemented two version with c/c++ and python, C++ version was implemented from scratch, pytho version designed with pyahocorasich package.For detail please check here

Document Classification

Document classification is a common application in task, I will write some implemenation with tensorflow.

• TextCNN was a multi-class classification with 7 labels, all the data was short titles.

Learning To Rank

Learning to rank is an important task in recommendation, here I collect the most common models and algorithms to reimplement them with real dataset. The different between the implementation with other repository in github was my implemenation can be used in your product project directly after modified the input pipeline proprecess.

DeepFM

• Add description
• Add impemenation

Contact

If you have any ideas or suggestion contact me with [email protected]