Twitter users spend an average of 4 minutes on social media Twitter. On an average of 1 minute, they read the same stuff. It shows that users spend around 25% of their time reading the same stuff.
Also, most of the tweets will not appear on your dashboard. You may get to know the trending topics, but you miss not trending topics. In trending topics, you might only read the top 5 tweets and their comments.
So, what are you going to do to avoid wastage of time on Twitter?
I would say summarize your whole trending Twitter tags data. And, then you can finish reading all trending tweets in less than 2 minutes.
In this article, I will explain to you how you can leverage Natural Language Processing (NLP) pre-trained models to summarize twitter posts based on hashtags. We will use 4 ( T5, BART, GPT-2, XLNet) pre-trained models for this job.
Each pre-trained model has its own architecture and weights. So, the summarization output given by these models could be different from each other.
Test the twitter data on different models and then choose the model which shows summarization close to your understanding. And then deploy that model into production.
Let’s start with collecting Twitter Live data.
You can get Twitter live data in 2 ways.
I will be using step 1 to fetch the data. Once you receive the credentials for Twitter API, follow the below code to get Twitter data through API.
Now, let’s start summarizing data using pre-trained models one by one.
T5 is a state of the art model used in various NLP tasks that includes summarization. We will be using the transformers library to download the T5 pre-trained model and load that model in a code.
The Transformers library is developed and maintained by the Hugging Face team. It’s an open-source library.
Know more about the T5 model here.
Here is code to summarize the Twitter dataset using the T5 model.
Observation on Code
BART uses both BERT (bidirectional encoder) and GPT (left to the right decoder) architecture with seq2seq translation. BART achieves the state of the art results in the summarization task.
BART pre-trained model is trained on CNN/Daily mail data for the summarization task, but it will also give good results for the Twitter dataset.
We will take advantage of the hugging face transformer library to download the T5 model and then load the model in a code.
Here is code to summarize the Twitter dataset using the BART model.
Observation on Code
GPT-2 model with 1.5 million parameters is a large transformer-based language model. It’s trained for predicting the next word. So, we can use this specialty to summarize Twitter data.
GPT-2 models come with various versions. And, each version’s size is more than 1 GB.
We will be using the bert-extractive-summarizer library to download GPT-2 models. Learn more about the bert-extractive-summarizer library here.
Use pip install bert-extractive-summarizer command to install the library.
Here is a code to summarize the Twitter dataset using the GPT-2 model.
Observation on Code
XLNet is an improved version of the BERT model which implement permutation language modeling in its architecture. Also, XLNet is a bidirectional transformer where the next tokens are predicted in random order.
The XLNet model has two versions xlnet-base-cased and xlnet-large-cased.
Here is a code to summarize the Twitter dataset using the XLNet model.
Observation on Code
Other use-cases of Summarization
In this article, we have summarized the Twitter live data using T5, BART, GPT-2, and XLNet pre-trained models. Each model generates a different summarize output for the same dataset. Summarization by the T5 model and BART has outperformed the GPT-2 and XLNet models.
These pre-trained models can also summarize articles, e-books, blogs with human-level performance. In the future, you can see a lot of improvements in summarization tasks. And this will help you to solve many summarization related tasks.