Deploy your apps to a supercloud in a few clicks

This Engineering Education program is supported by Section. Instantly deploy your GitHub apps, Docker containers or K8s namespaces to a supercloud.

Try It For Free

How to Build an NLP Based Emotion Detection Model using Neattext and Scikit-learn

November 30, 2021

Natural language processing helps computers understand speech and written text like a human being. This allows machines to compute necessary responses. One of the many NLP applications is emotion detection in text.

The emotion detection model is a type of model that is used to detect the type of feeling and attitude in a given text. It may be a feeling of joy, sadness, fear, anger, surprise, disgust, or shame.

An emotion detection model can classify a text into the following categories. By using emotion detection in text, businesses can know how customers feel about their brand and products. This helps businesses improve product quality and service delivery.

In this tutorial, we will use Neattext and Scikit-learn to build our model. Neattext is a Python library used to preprocess our dataset. Neattext will clean the text dataset by removing stop words and other noise.

This makes it easier for the model to use the dataset during training. We’ll use Scikit-learn to build our model. It contains all the algorithms required for classification. This is a practical guide from data preprocessing to model building and testing.


To follow along a reader must know:

NOTE: In this tutorial, we will use Google Colab to build our model.

Table of contents

Exploring our dataset

In this tutorial, we will use a dataset that contains various texts with different emotion labels. The dataset has eight emotion labels that are named as follows: joy, sadness, fear, anger, surprise, neutral, disgust, and shame. The dataset will be used during the training phase. The trained model will then be used to classify a given text into the emotion labels.

A snip of the dataset is shown below:

Dataset snip

To get this dataset in CSV format, click here.

After downloading the dataset, we can now load it into our Google Colab notebook.

Loading exploratory data analysis packages

We import two exploratory data analysis packages, Pandas and Numpy. We will use pandas to read our CSV file and load it into our Google Colab notebook. Numpy works with arrays and is used to perform mathematical computations to our data.

import pandas as pd
import numpy as np

We use pandas to load our dataset.

Loading dataset

Use the following command to load the dataset:

df = pd.read_csv("emotion-dataset.csv")

To see how the dataset is structured, use this command:


The output is shown below:

Dataset structure

The image above shows that our dataset has two columns: Emotion and Text. The Emotion column represents the various emotion labels. The Text column shows all the texts in our dataset.

To show the value count for each emotion, run the code below. It will give you the total number of texts for each emotion label.


The output is shown below:

Value count

We can now start cleaning our dataset using Neattext.

Getting started with Neattext

As mentioned earlier, Neattext is a Python library that is used to preprocess our dataset. Neattext will clean the text dataset by removing stop words and noise.

To install Neattext, run this command:

!pip install neattext

We import neattext as follows:

import neattext.functions as nfx

To use neattext, we list all the methods and attributes used by neattext for data cleaning.


The output for the methods and attributes is shown below:

Methods and attributes

We are interested in only two methods from the list, the remove_stopwords and remove_userhandles.

Removing user handles

The dataset contains some Twitter handles of different users. These handles are unnecessary and considered as noise to our dataset. We use the remove_userhandles method to remove them from our dataset.

df['Clean_Text'] = df['Text'].apply(nfx.remove_userhandles)

We use the apply() method to add remove_userhandles. We save the cleaned dataset into a new column named Clean_Text.

Removing stopwords

Stopwords is a list of all the commonly used words in any language. Stopwords carry very little helpful information and have minimal impact on the model during training. These words lead to model bias during training. Removing stopwords eliminates unimportant words, allowing the applications to focus on the essential words instead.

Common stopwords are like articles of a given language. They include the words, the, is, and and are in the English language.

df['Clean_Text'] = df['Clean_Text'].apply(nfx.remove_stopwords)

In this section, we also use apply() to add remove_stopwords. We save the cleaned dataset into a new column named Clean_Text.

To get the output of the clean dataset, run this command:


The output of the dataset after removing user handles and stopwords is shown below:

Clean dataset

Now that we cleaned our dataset, we can now load our machine learning packages.

Importing machine learning packages

from sklearn.linear_model import LogisticRegression
from sklearn.feature_extraction.text import CountVectorizer
from sklearn.model_selection import train_test_split
from sklearn.metrics import accuracy_score

From the code above, we have imported the following:

  • LogisticRegression is an algorithm used for both classification and regression. This algorithm is imported from Scikit-learn. We will use it for emotion classification.

  • Machine learning models have a problem comprehending raw text, they work well with numbers. Machines cannot process the raw text data, and it has to be converted into a matrix of numbers. CountVectorizer is used to convert the raw text into a matrix of numbers. This process depends on the frequency of each word in the entire text. During this process, CountVectorizer extracts important features from the text. They are used as input for the model during training.

  • The train_test_split method is important during the splitting of the dataset. It splits the dataset set into two sets, a train set, and a test set. This depends on the percentage specified by the user.

  • The accuracy_score is important when calculating the accuracy score of our model during prediction.

Features and labels are important in machine learning. In this section, we will specify our features and labels.

Model features and labels

Features are the attributes and variables extracted from the dataset. These extracted features are used as inputs to the model during training enabling model learning. Our features are present in the Clean_Text column.

Labels are the output or the target variable. Our label is the Emotion column, and this is what the model is predicting.

Xfeatures = df['Clean_Text']
ylabels = df['Emotion']

Dataset splitting

We need to split our dataset into a train set and test set. The model will learn from the train set. We will use the test set to evaluate the model performance and measure the model’s knowledge capability.

We specify the test_size=0.3. This will split our dataset with 70% of data used for training and 30% for testing.

To make the process of training our model faster and automated, we will use a machine learning pipeline. Machine learning pipelines automate the machine learning workflows such as model fitting and training.

Doing so saves the developers time and reduces model errors. Machine learning pipeline is a new approach adopted by many developers as it produces quality models which are free from bugs.

To use this pipeline approach, we need to import the Pipeline package.

Pipeline approach

We import Pipeline using the following code:

from sklearn.pipeline import Pipeline

To use Pipeline we need to specify the machine learning stages we want to automate. In this tutorial, we have two processes we want to automate. The first stage is the CountVectorizer process. This stage converts the raw text dataset into a matrix of numbers that a machine can understand.

The second stage is the model training process using the LogisticRegression algorithm. In this stage, the model learns from the dataset. During training, it understands patterns, gains knowledge, and uses the knowledge to make predictions.

The two pipeline stages are initialized as follows:

pipe_lr = Pipeline(steps=[('cv',CountVectorizer()),('lr',LogisticRegression())])

After initializing the two stages, we need to fit these stages into our dataset. We will use the train set dataset, which is specified as x_train and y_train.

Model fitting

To fit the pipeline stages into x_train and y_train, run this code:,y_train)

The Pipeline will run the following stages automatically and produce the following output:

Pipeline ouput

This process produces the optimal model that will give the best results. With time the model will improve on its own and give better prediction results.

After building the model, we check the accuracy score produced by our Pipeline to evaluate the model’s performance.

Calculating the accuracy score

To check the accuracy score, run this command:


The output is shown below:


When the accuracy score is expressed as a percentage, it becomes 82.0%. This is a high accuracy after the first phase of training. Through continuous training, the model will increase the accuracy score. The higher the accuracy score, the better model will be in making predictions.

Our model is now fully trained and tested. We can now use this model to make predictions.

Making a single prediction

Our model should be able to classify a given text into emotion labels. Let’s use the sample text shown below to make a prediction.

sample1 = "This chocolate was very sweet it made me happy"

To make the actual prediction, run this code:


We use the predict method to predict our sample text. The prediction outcome is shown below:

array(['joy'], dtype=object)

The prediction outcome is joy. This is the correct prediction. This shows that our model can accurately predict. It can be adopted and deployed to production.

The Google Colab notebook link for this tutorial is available here.


In this tutorial, we have learned how to build an emotion detection model using Neattext and Scikit-learn. We started by cleaning our dataset using Neattext. The dataset has to be in the right shape before it is used for training.

We explored the different machine learning packages and algorithms to use. We used the logistic regression algorithm to build our emotion detection model. We also introduced a concept known as machine learning pipeline. The pipeline approach made our work easier. It automates the CounterVectorizer process and model building.

Finally, we used this model to make accurate predictions. Any business can adopt the model and use it to monitor online conversations and reviews made by customers. This gives businesses an added advantage on how to improve their brand image.

Happy coding!


Peer Review Contributions by: Willies Ogola