ICT Companies Social Network Analysis

Natural Language Processing or NLP is the field of study on computational analysis of human language. This area of knowledge includes a very wide variety of techniques and applications. One of them, within the field of language analysis and comprehension, is Sentiment Analysis, an application that allows a text to be classified according to its positive, negative or neutral charge or polarity.

In this post, with a few lines of python code we’ll do the following tasks:

• Connect to the Twitter API
• Download the latest tweets that mention certain ICT companies
• Use a pre-trained machine learning model to perform sentiment analysis of these tweets
• Visualize the results

The pre-trained model that we are going to use is RoBERTuito, a model trained with 500 million tweets in Spanish. The authors of the paper/model made it available through the platform HuggingFace and the library pysentimento to facilitate NLP research and applications in Spanish.

Clarification 1: It is natural and expected that mentions of ICT companies in social media have a negative sentiment, since it is one of the channels for submitting complaints and, as it is a paid service, it is unusual to post a positive comment in case there are no problems with the service.

Clarification 2: to access the tweets, it is necessary to first apply for authentication credentials at Twitter for Developers. Once you have the credentials you should save them in a file called search_tweets_creds.yml with the following structure:

search_tweets_api:
    bearer_token: MY_BEARER_TOKEN
    endpoint: https://api.twitter.com/2/tweets/search/recent

To obtain the tweets I will use the searchtweets-v2 library, a Python Client for the Twitter API Version 2.

Use the following code for authentication and to obtain the last 100 tweets that mention each of the companies of interest:

from searchtweets import load_credentials, ResultStream, gen_request_parameters, collect_results

search_args = load_credentials(filename="search_tweets_creds.yml", 
                               yaml_key="search_tweets_api",
                               env_overwrite=False)

empresas = ["Telecentro", "MovistarArg", "ClaroArgentina", "PersonalAr"]
empresas_tweets = dict()

for empresa in empresas:
    query = gen_request_parameters(empresa, results_per_call=100, granularity=None)
    tweets = collect_results(query,
                             max_tweets=100,
                             result_stream_args=search_args)
    empresas_tweets[empresa] = tweets[0]['data']

Pre-process tweets, apply sentiment analysis and extract the category for each of the tweets and companies:

from pysentimiento import create_analyzer

analyzer = create_analyzer(task="sentiment", lang="es", model_name="pysentimiento/robertuito-sentiment-analysis")

empresas_tweets_sent = dict()
empresas_tweets_sent_out = dict()

for empresa in empresas:
    empresas_tweets_sent[empresa] = [analyzer.predict(tuit) for tuit in empresas_tweets_proc[empresa]]
    empresas_tweets_sent_out[empresa] = [tuit.output for tuit in empresas_tweets_sent[empresa]]

Visualize the results:

import numpy as np
import matplotlib.pyplot as plt

empresas_tweets_sent_count = dict()
fig, axes = plt.subplots(2, 2, figsize=(8, 6),dpi=144)

plt.suptitle("Análisis de Sentimientos de Empresas TIC")

array_index = [(0,0), (0,1), (1,0), (1,1)]
axes_title_font_size = 10

for empresa, index in zip(empresas, array_index):
    empresas_tweets_sent_count[empresa] = np.unique(empresas_tweets_sent_out[empresa], return_counts=True)
    axes[index].pie(empresas_tweets_sent_count[empresa][1], labels=empresas_tweets_sent_count[empresa][0], wedgeprops=dict(width=.5), autopct='%1.f%%')
    axes[index].set_title(empresa, fontsize=axes_title_font_size)