A Generic Sentence Embedding Library

In natural language processing, we need to encode text data. In the past, we mostly use encoders such as one-hot, term frequency, or TF-IDF (normalized term frequency). There are many challenges with these techniques. In the recent years, the latest advancements give us opportunity to encode sentences or words in more meaningful formats. The word2vec technique and BERT language model are two important ones.

The sentence embedding is an important step of many NLP projects from sentiment analysis to summarization. We believe that a flexible sentence embedding library is needed to build prototypes fast. That is why we have initiated this project. In the early releases, you will have access to the standard encoders. We will add more curated techniques in the later releases. Hope you can use this library in your exciting NLP projects.

Library

The package requires the following libraries:

• gensim
• numpy
• spacy
• transformers
• torch

The sent2vec package is developed to help you prototype faster. That is why it has many dependencies on other libraries.

Install

It can be installed using pip:

pip3 install sent2vec

Usage

If you want to use the the BERT language model (more specifically, distilbert-base-uncased) to encode sentences for downstream applications, you must use the code below. Now, you can compute distance among sentences by using their representational vectors. In the example, as expected, the distance between vectors[0] and vectors[1] is less than the distance between vectors[0] and vectors[2].

from sent2vec.vectorizer import Vectorizer

sentences = [
    "This is an awesome book to learn NLP.",
    "DistilBERT is an amazing NLP model.",
    "We can interchangeably use embedding, encoding, or vectorizing.",
]
vectorizer = Vectorizer()
vectors = vectorizer.bert(sentences)

dist_1 = cosine_distance(vectors[0], vectors[1])
dist_2 = cosine_distance(vectors[0], vectors[2])

print('dist_1: {}'.format(dist_1), 'dist_2: {}'.format(dist_2))
dist_1: 0.043, dist_2: 0.192

If you want to use a word2vec approach instead, you must first split sentences to lists of words using the sent2words method. In this stage, you can customized the list of stop-words by adding or removing to/from the default list. When you extract the most important words in sentences, you can compute the sentence embeddings using the w2v method. This method computes the average of vectors corresponding to the remaining words using the code bleow.

from sent2vec.vectorizer import Vectorizer

sentences = [
    "Alice is in the Wonderland.",
    "Alice is not in the Wonderland.",
]
vectorizer = Vectorizer()
words = vectorizer.sent2words(sentences, remove_stop_words=['not'], add_stop_words=[])
model_path = os.path.join(os.path.abspath(os.getcwd()), 'glove-wiki-gigaword-300')
vectors = vectorizer.w2v(words, model_path= model_path)

print(words)
[['alice', 'wonderland'], ['alice', 'not', 'wonderland']]

As seen above, you can use different word2ved model by sending its path to the w2v method. You can use a pre-trained model or a customized one.

And, that's pretty much it!