First of all we recommend you to install the requirements.

!pip3 install -r requirements.txt

Also, it is necessary to install some spacy models:

!python -m spacy download en_core_web_sm
!python -m spacy download es_dep_news_trf

Data cleaning is the process of preparing data for analysis by removing or modifying data that is incorrect, incomplete, irrelevant, duplicated, or improperly formatted. In this file there are two scripts: NLPSpanish_DataCleaning.ipynb and data_cleaning.py.

• Exploring the dataset
• Removing punctuation
• Removing line breaks and extra whitespaces
• Normalize text
• Removing accents
• Stemming
• Lemmatization
• Removing stopwords
• Removing twitter nicknames, emails and numbers
• Word and Sentence Tokenization

The last script is the method with all the functions defined. Here is a brief example of how to use it:

import data_cleaning.spanish as sp

sentence = 'La cordillera es un sector de latinoamérica muy lindo. Tiene mucha naturaleza,\
  animales y paisajes. ¡Quiero visitarla pronto, antes de tener que irme a vivir a otro lugar! :('

cleaned_sent, entity = sp.text_preprocessing(sentence)
print(f'cleaned sentence: {cleaned_sent}')
# → cleaned sentence: ['cordillerar', 'sector', 'latinoamerico', 'lindo', 'mucho', 'naturaleza', 'animal', 'paisaje', 'querer', 'visitar','él','pronto','tener','ir','yo','vivir', 'lugar']

print(f'entity detected: {entity}')
# → entity detected: [('cordillera sector latinoamerica', 'LOC')]

Also you can work with corpus i.e. with tokenized sentences

document = sp.sentenceTokenize(sentence)
print(f'corpus: {document}')
# → corpus: ['La cordillera es un sector de latinoamérica muy lindo.', 'Tiene mucha naturaleza, animales y paisajes.', '¡Quiero visitarla pronto, antes de tener que irme a vivir a otro lugar!', ':(']

cleaned_doc,entities = sp.corpus(dataset=document)
print(f'cleaned corpus: {cleaned_doc}')
# → cleaned corpus: [['cordillerar', 'sector', 'latinoamerico', 'lindo'], ['mucho', 'naturaleza', 'animal', 'paisaje'], ['querer', 'visitar', 'él', 'pronto', 'tener', 'ir', 'yo', 'vivir', 'lugar']]
print(f'entities detected: {entities}')
# → entities detected: [[('cordillera sector latinoamerica lindo', 'LOC')]]

(back to top)

Working with machine learning models means that you must expressed numerically the information that you want to process. Therefore, there are many algorithms that enable words to be expressed mathematically, such as Bag-Of-Word, TF-IDF, Word2Vec, FastText. This is called feature extraction or feature encoding. In this archive you will find a brief introduction to feature encoding methods using a spanish twitter dataset.

With w2v.ipynb you will be able to understand Skip-gram and Continuous Bag-of-Words method. Also an example of how to train a model and how to use a pretrained W2V.

This word embedding uses a Skip-gram method, in fastText.ipynb you can find a training example and a pretrained model.

GloVe.ipynb has a spanish pretrained model example.

(back to top)

Topic modeling is a NLP task that consists in recognizing the words from the topics present in the documents. In this section we use only unsupervised methods, there are two scripts: topic_modeling.ipynb and topic_m.py. The first script is a tutorial of this task.

• Bar chart word count and weights of each topic keyword
• A fast pyLDAvis visualization
• Dataframe with the ppl vocabulary and score per topic
• Dataframe with the dominant topic and its percentage contribution in each document
• Dataframe with the most representative sentence for each topic
• Sentence chart colored by topic
• TSNE interactive topic clustering

Here is an example of how to use the script and the expected output:

import topic_m as tm

model_list, coherence_values = tm.compute_coherence_values(dictionary=id2word, corpus=corpus, id2word=id2word, texts=cleaned_doc_train, start=1, limit=10, step=2, model_='LDA')
lda_model = model_list[coherence_values.index(max(coherence_values))]
tm.tsne_plot(model_=lda_model,corp=corpus,name_model='LDA',text=cleaned_doc_train,interactive_labels = True)

(back to top)

Text classification is the process of categorizing the text into a group of words. In the following scripts were tested machine and deep learning models such as LSTM, Naive Bayes, Transformers, Random Forest, etc. The Notebook TC_complaints_classifiers.ipynb looks for the best model (using cross-validation methods) for a translated Complains Dataset.

An especific case of text classification is sentiment analysis. This focuses on the polarity of a text (positive, negative, neutral). For TC_sentiment_analysis.ipynb we used a Hate Speech Spanish Dataset. Models trained:

• LSTM
• Support-vector machin
• Random Forest Classifier
• Logistic Regression
• Naive Bayes
• Finetuned Transformer

(back to top)

In ner.ipynb we show you two differents ways for recognizing entities.

• Using a pretrained transformer from Hugging Face
• Using the data_cleaning.ipynb functions There are not trained models because NER's dataset are expensive.

(back to top)

We explored some ideas with the WhatsApp Dataset.

Testing LSI and LSA, in topic_modeling_wsp.ipynb we extracted the topics of the most popular group chat.

In word_embeddings_keybert.ipynb there are two word embeddings trained with the wsp messages; Word2Vec and FastText. Both were tested with the KeyBert arquitecture.

There are just 550 labeled rows even though it is a small amount of labeled data, we experimented with zero-shot learning trained transformers and then with multiclass classification methods. Hopefully, in some future we expect to be able to guide the data labeling process.

(back to top)

We search and experiment with some autocorrect techniques. autocorrect_datasetwsp.ipynb contains a autocorrect method using FastText. Also, the palabras_chilenas.csv and corrected_with_speller.csv datasets were created.