Classificação e agrupamento de crimes relacionados à violência contra mulheres : um estudo baseado em Doc2Vec

Abstract

Word embedding is a technique that enables the transformation of words into continuous vector representations, allowing machine learning algorithms to process linguistic information more efficiently. Doc2vec is a Natural Language Processing (NLP) model developed as an extension of the word2vec algorithm, which creates word embeddings using two architectures: continuous bag of words (CBOW) and continuous skip-gram. CBOW is based on the conditional probability of generating a central word from context words, while skip-gram is based on the conditional probability of generating context words from a central word. While word2vec creates continuous vectors only for words, doc2vec uses two architectures to create its vectors: PV-DM (Distributed Memory Model of Paragraph Vectors), which creates vectors for entire documents, such as paragraphs, sentences, or even larger text blocks, capturing the syntactic, semantic, and contextual relationships associated with documents that belong to a text collection. The other technique is called PV-DBOW (Distributed Bag of Words of Paragraph Vectors), which is similar to the skip-gram model in word2vec as it ignores context words and forces the model to predict words from the entire document, sampled randomly, at the output. Using the PV-DM architecture, a doc2vec model was trained to create continuous vectors from crime report narratives related to violence against women for document classification and clustering, detection of sentences revealing the most prevalent types of violence, and the associations between words, sentences, and documents to better identify crimes.