nlpipes 0.1.18

Text Classification with Transformers

Text Classification with Transformers

Documentation • References

Overview

NLPipes is for people unfamiliar with Transformers who want an end to end solution to solve practical text classification problems, including:

• Single-label classification: A typical use case is sentiment detection where one want to detect the overall sentiment polarity (e.g., positive, neutral, negative) in a review.
• Multi-label classification: A typical use case is aspect categories detection where one want to detect the multiple aspects mentionned in a review (e.g., product_quality, delivery_time, price, ...).
• Aspect-based classification: A typical use case is aspect based sentiment analysis where one want to detect sentiment polarity associated to each aspect categories mentionned in a review (e.g., product_quality: neutral, delivery_time: negative, price: positive, ...).

NLPipes expose a Model API that provide a unique and simple abstraction for all the tasks. The library maintain a common usage pattern across models (train, evaluate, predict, save) with also a clear and consistent data structure (python lists as inputs/outputs data).

Built with

NLPipes is built with TensorFlow and HuggingFace Transformers:

• TensorFlow: An end-to-end open source deep learning framework
• Transformers: An general-purpose open-sources library for transformers-based architectures

Getting Started

Installation

1. Create a virtual environment

python3 -m venv nlpipesenv
source nlpipesenv/bin/activate

2. Install the package

pip install nlpipes

Tasks

A model can be trained for a specific task by first loading a backbone model. The train command takes at minimum two parameters (X and Y), where X is a list of texts to train on and Y is the training target.

The training target expect different formats, depending on what task you want to solve:

Single Label Classification:

Give one label name for each sequence of text in X:

 model = Model("albert-base-v2",
               task='single-label-classification',
               all_labels=["NEG", "NEU", "POS"],
              )
 
 X = ["This was bad.", "This was great!"]
 Y = ["NEG", "POS"]
 
 model.train(X, Y)

Multiple Label Classification:

Give a list of class names for each sequence of text in X:

 model = Model("albert-base-v2",
               task='multi-label-classification',
               all_labels=all_labels,
               )
 
 X = ["I want a refund!",
      "The bill I got is not correct and I also have technical issues",
      "All good"]
 Y = [
      ["billing"], 
      ["billing", "tech support"],
      []
     ]
     
 model.train(X, Y)

Aspect Based Classification:

Give a list of lists of label lists (pairs) for each given text in X:

 model = Model("albert-base-v2",
               task='class-label-classification',
               all_labels=["NEG", "NEU", "POS"],
              )
 
 X = ["The room was nice.",
      "The food was great, but the staff was unfriendly.",
      "The room was horrible, but the waiters were welcoming"]
 Y = [
      [["room", "POS"],
      [["food", "POS"], ["staff", "NEG"]],
      [["room", "NEG"], ["staff", "POS"]],
     ]
     
 model.train(X, Y)

Examples

Here are some examples on open datasets that show how to use NLPipes on different tasks:

Name	Notebook	Description	Task	Size	Memory	Speed
GooglePlay Sentiment Detection	Available	Train a model to detect the sentiment polarity from the GooglePlay store	Single label classification
StackOverflow tags Detection	Available	Train a model to detect tags from the StackOverFlow questions	Multiple label classification
Amazon Aspect Based Sentiment Detection	Available	Train a model to detect the aspect based sentiment polarity on Laptops Amazon reviews	Class label classification

Notices

• NLPipes is still in its early stage. The library comes with no warranty and future releases could bring substantial API and behavior changes.