NLP-Classification-MCTI

MCTI Text Classification Task (uncased)

Disclaimer: The Brazilian Ministry of Science, Technology, and Innovation (MCTI) has partially supported this project.

The model NLP MCTI Classification Multi is part of the project Research Financing Product Portfolio (FPP) and focuses on the Text Classification task, exploring different machine learning strategies to classify a small amount of long, unstructured, and uneven data to find a proper method with good performance. Pre-training and word embedding solutions were used to learn word relationships from other datasets with considerable similarity and a larger scale. Then, using the acquired resources, based on the dataset available in the MCTI, transfer learning plus deep learning models were applied to improve the understanding of each sentence.

According to the abstract,

"The research results serve as a successful case of artificial intelligence in a federal government application". More details about the project, architecture model, training model, and classifications process can be found in the article "Using transfer learning to classify long unstructured texts with small amounts of labeled data".

Model description

The work consists of a machine learning model with word embedding and Convolutional Neural Network (CNN). For the project, a Convolutional Neural Network (CNN) was chosen, as it presents better accuracy in empirical comparison with 3 other different architectures: Neural Network (NN), Deep Neural Network (DNN), and Long-Term Memory (LSTM).

As the input data is composed of unstructured and nonuniform texts, it is essential to normalize the data to study little insights and valuable relationships to work with their best features. In this way, learning is facilitated and allows the gradient descent to converge more quickly.

The first layer of the model is a pre-trained Word2Vec embedding layer as a method of extracting features from the data that can replace one-hot coding with dimensional reduction. The pre-training of this model is explained further in this document.

After the embedding layer, there is the CNN classification model. The architecture of the CNN network is composed of a 50% dropout layer followed by two 1D convolution layers associated with a MaxPooling layer. After maximum grouping, a dense layer of size 128 is added and connected to a 50% dropout which finally connects to a flattened layer and the final sort dense layer. The dropout layers helped avoid network overfitting by masking part of the data so that the network learned to create redundancies in analyzing the inputs.

Model variations

Table 1 below presents the results of several implementations with different architectures, highlighting the accuracy, f1-score, recall, and precision results obtained in each network training.

Table 1: Results of experiments

Table 2 below shows the times required for training each epoch, the data validation execution time and the weight of the deep learning model associated with each implementation.

Table 2: Results of Training time epoch, Validation time, and Weight

In addition, it is possible to notice that the model of Longformer + SNN and Longformer + LSTM were not able to learn. Perhaps the models need some adjustments; however, each training attempt takes between 5 and 8 hours, which made an attempt to adjust unfeasible because of other models already showing promising results.

With Longformer, the problems caused by the size the model's size became more visible. First, it was necessary to actively deallocate unused chunks of memory right after use so that the next steps could be loaded. Then, it was necessary to use a CPU environment for training the networks because the model's weight exceeded the 16GB of video memory available on the P100 board, available in Colab during training. In this case, the high RAM environment was used, which delivers 25GB of memory for use with the CPU, and this means a longer time is required for training since the GPU performs matrix operations faster than the CPU. These models were trained 5x with 100 training epochs each.

Intended uses

How to use

This model is available in Hugging Face spaces to be applied to excel files containing scrapped opportunity data.

• NLP MCTI Classification Multi

The training and evaluation notebooks can be found in the github repository:

• PPF-MCTI Repository

Limitations and bias

This model is uncased: it does not make a difference between english and English.

This model depends on high-quality scrapped data. Since the model understands a finite number of words, the input needs to have little to no wrong encodings and abstract markdowns so that the preprocessing can remove them and correctly identify the words.

Even if the training data used for this model could be characterized as fairly neutral, this model can have biased predictions:

Performance limiting: Loading the longformer model in the memory means needing 11Gb available only for the model without considering the weight of the deep learning network. For training, this means we need a 20+ Gb GPU to perform the training. Here this was resolved using the high RAM environment of google Colab Pro and training using CPU, which justifies the longer training time per season.

Replicability limitation: Due to the simplicity of the Keras embedding model, we are using one-hot encoding, and it has a delicate problem for replication in production. This detail is pending further study to define whether it is possible to use one of these models.

This bias will also affect all fine-tuned versions of this model.

Training data

The inputted training data was obtained from scrapping techniques over 30 different platforms, e.g., The Royal Society, Annenberg foundation, and contained 928 labeled entries (928 rows x 21 columns). Of the data gathered, was used only the main text content (column u). Text content averages 800 tokens in length, but with high variance, up to 5,000 tokens.

Training procedure

Model training with Word2Vec embeddings

After the pre-trained model of word2vec embeddings had already learned meanings relevant to the classification problem, it was coupled to the classification model to train it with the labeled data in a supervised way. Table 6 shows the results obtained with related metrics. This implementation reached new levels of accuracy, with 86% for CNN architecture and 88% for the LSTM architecture.

Preprocessing

Pre-processing was used to standardize the texts for the English language, reduce the number of insignificant tokens and optimize the training of the models.

The following assumptions were considered:

• The Data Entry base is obtained from the result of Goal 4.
• Labeling (Goal 4) is considered true for accuracy measurement purposes;
• Preprocessing experiments compare accuracy in a shallow neural network (SNN);
• Pre-processing was investigated for the classification goal.

From the Database obtained in Goal 4, stored in the project's GitHub, a Notebook was developed in Google Colab to implement the preprocessing code, which also can be found on the project's GitHub.

Several Python packages were used to develop the preprocessing code:

Table 3: Python packages used

As detailed in the notebook on GitHub, in the pre-processing, code was created to build and evaluate 8 (eight) different bases, derived from the base of goal 4, with the application of the methods shown in table 4.

Table 4: Preprocessing methods evaluated

First, the treatment of punctuation and capitalization was evaluated. This phase resulted in the construction and evaluation of the first four bases (xp1, xp2, xp3, xp4).

Then, the content simplification was evaluated, from the xp4 base, considering stemming (xp5), Lemmatization (xp6), stemming + stopwords removal (xp7), and Lemmatization + stopwords removal (xp8).

All eight bases were evaluated to classify the eligibility of the opportunity through the training of a shallow neural network (SNN – Shallow Neural Network). The metrics for the eight bases were evaluated. The results are shown in Table 5.

Table 5: Results obtained in Preprocessing

Even so, between these two excellent options, one can judge which one to choose. XP7: It has less training time, and less number of unique tokens. XP8: It has smaller maximum sizes. In this case, the criterion used for the choice was the computational cost required to train the vector representation models (word-embedding, sentence-embeddings, document-embedding). The training time is so close that it did not have such a large weight for the analysis.

As the last step, a spreadsheet was generated for the model (xp8) with the fields opo_pre and opo_pre_tkn, containing the preprocessed text in sentence format and tokens, respectively. This database was made available on the project's GitHub with the inclusion of columns opo_pre (text) and opo_pre_tkn (tokenized).

Pretraining

Since labeled data is scarce, word-embeddings were trained in an unsupervised manner using other datasets that contain most of the words it needs to learn. The idea implemented was based on introducing better and better-trained word embeddings in the model. For an additional dataset to be applied to improve word-embedding training, it must be compatible with the dataset used to train the classifier. We searched for datasets from the Kaggle, a platform with over a thousand available NLP datasets, and the closest we found was the BBC News Articles dataset, which achieved only 56% compatibility.

The alternative was to use web scraping algorithms to acquire more unlabeled data from the same sources, thus ensuring compatibility. The original dataset had 260 labeled entries.

Table 6: Compatibility results (*base = labeled MCTI dataset entries)

Table 7: Results from Pre-trained WE + ML models

Evaluation results

The table below presents the results of accuracy, f1-score, recall, and precision obtained in the training of each network. In addition, the necessary times for training each epoch, the data validation execution time, and the weight of the deep learning model associated with each implementation were added.

Table 8: Results of experiments

The results obtained surpassed those achieved in goal 6 and goal 9, with the best accuracy obtained of 94% in the longformer + CNN model. We can also observe that the models that achieved the best results were those that used the CNN network for deep learning.

With the motivation to increase accuracy obtained with baseline implementation, was implemented a transfer learning strategy under the assumption that small data available for training was insufficient for adequate embedding training. In this context, were considered two approaches:

• Pre-training word embeddings using similar datasets for text classification;
• Using transformers and attention mechanisms (Longformer) to create contextualized embeddings.

Templates using Word2Vec and Longformer also need to be loaded and their weights are as follows:

Table 9: Templates using Word2Vec and Longformer

In addition, it was possible to notice that the model of longformer + SNN and longformer + LSTM were not able to learn. Perhaps the models need some adjustments, but each training attempt took between 5 and 8 hours, which made it impossible to try to adjust when other models were already showing promising results.

Above the results obtained, it is also necessary to highlight two limitations found for the replication and training of networks:

These 10Gb of the model exceeded the Github limit and did not go to the repository, so to run the system, we need to download the pre-trained network in the notebook and run the encoder-decoder with the data to create the model. It is advisable to do this in a GPU environment and save the file on the drive. After that, change the environment to CPU to perform the training. Trying to generate the model unsing the CPU will take more than 3 hours of processing.

The best model that does not have any limitations is Word2Vec + CNN. However, we need to study the limitations to understand whether it is possible to introduce a new model with better accuracy and indicators. These adjustments will be worked on during goals 13 and 14, where the main objective will be to encapsulate the solution in the most suitable way for use in production.

Benchmarks

BibTeX entry and citation info

@conference{webist22,
author       ={Carlos Rocha. and Marcos Dib. and Li Weigang. and Andrea Nunes. and Allan Faria. and Daniel Cajueiro.
               and Maísa {Kely de Melo}. and Victor Celestino.},
title        ={Using Transfer Learning To Classify Long Unstructured Texts with Small Amounts of Labeled Data},
booktitle    ={Proceedings of the 18th International Conference on Web Information Systems and Technologies - WEBIST,},
year         ={2022},
pages        ={201-213},
publisher    ={SciTePress},
organization ={INSTICC},
doi          ={10.5220/0011527700003318},
isbn         ={978-989-758-613-2},
issn         ={2184-3252},
}

<a href="https://huggingface.co/exbert/?model=bert-base-uncased"> <img width="300px" src="https://cdn-media.huggingface.co/exbert/button.png"> </a>