The article:

https://medium.com/@adriensieg/text-similarities-da019229c894

Methods studied in the article

• Jaccard Similarity

• Different embeddings+ K-means

• Different embeddings+ Cosine Similarity

• Word2Vec + Smooth Inverse Frequency + Cosine Similarity

• Different embeddings+LSI + Cosine Similarity

• Different embeddings+ LDA + Jensen-Shannon distance

• Different embeddings+ Word Mover Distance

• Different embeddings+ Variational Auto Encoder (VAE)

• Different embeddings+ Universal sentence encoder

• Different embeddings+ Siamese Manhattan LSTM

• Knowledge-based Measures

SOURCES

• [Incredible !!!] : https://github.com/nlptown/nlp-notebooks

  • An Introduction to Word Embeddings
  • Data exploration with sentence similarity
  • Discovering and Visualizing Topics in Texts with LDA (en français !)
  • Keras sentiment analysis with Elmo Embeddings
  • Multilingual Embeddings - 1. Introduction
  • Multilingual Embeddings - 2. Cross-lingual Sentence Similarity
  • Multilingual Embeddings - 3. Transfer Learning
  • NLP with pretrained models - spaCy and StanfordNLP
  • Named Entity Recognition with Conditional Random Fields
  • Sequence Labelling with a BiLSTM in PyTorch [sequence labelling tasks such as part-of-speech tagging or named entity recognition]
  • Simple Sentence Similarity [Word Mover’s Distance + Smooth Inverse Frequency + InferSent + Google Sentence Encoder + Pearson correlation]
  • Text classification with BERT in PyTorch
  • Text classification with a CNN in PyTorch
  • Traditional text classification with Scikit-learn [ELI5]
  • Updating spaCy’s Named Entity Recognition System

• [Tutorial WMD in jupyter notebook] : https://github.com/makcedward/nlp/blob/master/sample/nlp-word_mover_distance.ipynb

• [Word Mover Distance] : https://www.kaggle.com/ankitswarnkar/word-embedding-using-glove-vector

• [lstm-gru-sentiment-analysis] : https://github.com/javaidnabi31/Word-Embeddding-Sentiment-Classification

https://towardsdatascience.com/elmo-contextual-language-embedding-335de2268604

• [Learning Word Embedding (Mathematics)] https://lilianweng.github.io/lil-log/2017/10/15/learning-word-embedding.html

• [A Beginner’s Aha Moments for Word2Vec] : https://yidatao.github.io/2017-08-03/word2vec-aha/

• [Glove, Word2Vec, Fastext classes] : https://github.com/makcedward/nlp/blob/master/sample/nlp-word_embedding.ipynb

• [!!! Very nice tutorial about how word2vec works] : https://towardsdatascience.com/word2vec-made-easy-139a31a4b8ae

• [!!! An implementation guide to Word2Vec using NumPy and Google Sheets] : https://towardsdatascience.com/an-implementation-guide-to-word2vec-using-numpy-and-google-sheets-13445eebd281

• [WordRank embedding: “crowned” is most similar to “king”, not word2vec’s “Canute”] : https://rare-technologies.com/wordrank-embedding-crowned-is-most-similar-to-king-not-word2vecs-canute/

• [Nice !!! ]: https://github.com/JacopoMangiavacchi/SwiftNLC/tree/master/ModelNotebooks

  • Create Model With GloVe Embedding Bidirectional With Attention
  • Create Model With FastText Embedding
  • Create Model With GloVe Embedding
  • Create Model With GloVe Embedding Bidirectional
  • Create Model With NLTK Embedding
  • Create Model With NS Linguistic Tagger Embedding

• [Tutorial from ENSAE] : http://www.xavierdupre.fr/app/papierstat/helpsphinx/notebooks/text_sentiment_wordvec.html#les-donnees

• [CoreML with GloVe Word Embedding and Recursive Neural Network - nice tutorial] : https://medium.com/@JMangia/coreml-with-glove-word-embedding-and-recursive-neural-network-part-2-ab238ca90970

• [Big Benchmark] : http://nlp.town/blog/sentence-similarity/

  • Average W2V
  • Average W2V + Stopwords
  • Average W2V + TFIDF
  • Average W2V + TFIDF + Stopwords
  • Average Glove
  • Average Glove + Stopwords
  • Average Glove + TFIDF
  • Average Glove + TFIDF + Stopwords
  • W2V + WMD
  • W2V + Stopwords + WMD
  • Glove + WMD
  • Glove + Stopwords + WMD
  • Smooth Inverse Frequency + W2V
  • Smooth Inverse Frequency + Glove
  • InferSent (INF)
  • GSE (Google Sentence Encoder)

  InferSent (INF) = pre-trained encoder that was developed by Facebook Research. It is a BiLSTM with max pooling, trained on the SNLI dataset, 570k English sentence pairs labelled with one of three categories: entailment, contradiction or neutral.

  GSE (Google Sentence Encoder) = Google’s answer to Facebook’s InferSent. It comes in two forms:

  • an advanced model that takes the element-wise sum of the context-aware word representations produced by the encoding subgraph of a Transformer model
  • a simpler Deep Averaging Network (DAN) where input embeddings for words and bigrams are averaged together and passed through a feed-forward deep neural network.

  ====-> work with Pearson correlation

• [How to predict Quora Question Pairs using Siamese Manhattan LSTM] : https://medium.com/mlreview/implementing-malstm-on-kaggles-quora-question-pairs-competition-8b31b0b16a07

• [Latent Semantic Indexing (LSI) - An Example with mathematics] : http://www1.se.cuhk.edu.hk/~seem5680/lecture/LSI-Eg.pdf

• [Finding similar documents with Word2Vec and WMD] : https://markroxor.github.io/gensim/static/notebooks/WMD_tutorial.html

• [Cosine Similarity] : https://www.machinelearningplus.com/nlp/cosine-similarity/

• [Tutorial on LSI] : http://poloclub.gatech.edu/cse6242/2018spring/slides/CSE6242-820-TextAlgorithms.pdf

http://robotics.stanford.edu/~scohen/research/emdg/emdg.html#flow_eqw_notopt

http://robotics.stanford.edu/~rubner/slides/sld014.htm

http://jxieeducation.com/2016-06-13/Document-Similarity-With-Word-Movers-Distance/

• [Beyond Cosinu > Jensen-Shannon + Hypothesis Test] : http://stefansavev.com/blog/beyond-cosine-similarity/

• [Great ressources with MANY MANY notebooks] : https://www.renom.jp/index.html?c=tutorial

• [LE TRANSPORT OPTIMAL, COUTEAU SUISSE POUR LA DATA SCIENCE]: https://weave.eu/le-transport-optimal-un-couteau-suisse-pour-la-data-science/

• [BEST TUTORIAL variational-autoencoders] : https://www.jeremyjordan.me/variational-autoencoders/

• [BEST TUTORIAL : Earthmover Distance !!!!!] : https://jeremykun.com/2018/03/05/earthmover-distance/

Problem: Compute distance between points with uncertain locations (given by samples, or differing observations, or clusters).

• [Introduction to Wasserstein metric (earth mover’s distance) -> Mathematics]: https://yoo2080.wordpress.com/2015/04/09/introduction-to-wasserstein-metric-earth-movers-distance/

• [Word Mover’s distance calculation between word pairs of two documents] : https://stats.stackexchange.com/questions/303050/word-movers-distance-calculation-between-word-pairs-of-two-documents

• [WMD + Word2Vec] : https://github.com/stephenhky/PyWMD/blob/master/WordMoverDistanceDemo.ipynb

• [Books about Optimal Transport] : https://optimaltransport.github.io/pdf/ComputationalOT.pdf

• [NICE !!!!!! How Autoencoders work - Understanding the math and implementation] : https://www.kaggle.com/shivamb/how-autoencoders-work-intro-and-usecases

• [Word2Vec to convert each question into a semantic vector then stack a Siamese network to detect if the pair is duplicate] : http://www.erogol.com/duplicate-question-detection-deep-learning/

• [Amazing !!!] : https://github.com/makcedward/nlp

  • Distance Measurement:
    • Euclidean Distance, Cosine Similarity and Jaccard Similarity
    • Edit Distance + Levenshtein Distance
    • Word Moving Distance (WMD)
    • Supervised Word Moving Distance (S-WMD)
    • Manhattan LSTM

  • Text Representation:

    1. Traditional Method

    • Bag-of-words (BoW)
    • Latent Semantic Analysis (LSA) and Latent Dirichlet Allocation (LDA)

    2. Character Level

    • Character Embedding

    3. Word Level

    • Negative Sampling and Hierarchical Softmax
    • Word2Vec, GloVe, fastText
    • Contextualized Word Vectors (CoVe)
    • Embeddings from Language Models (ELMo)
    • Generative Pre-Training (GPT)
    • Contextual String Embeddings
    • Self-Governing Neural Networks (SGNN)
    • Multi-Task Deep Neural Networks (MT-DNN)
    • Generative Pre-Training-2 (GPT-2)
    • Universal Language Model Fine-tuning (ULMFiT)

    4. Sentence Level

    • Skip-thoughts
    • InferSent
    • Quick-Thoughts
    • General Purpose Sentence (GenSen)
    • Bidirectional Encoder Representations from Transformers (BERT)

[Zoom] : Google Sentence Encoder

• [Reference] : https://ai.googleblog.com/2018/05/advances-in-semantic-textual-similarity.html

• [Waaaaaaaaaaaa] :https://machinelearningmastery.com/encoder-decoder-deep-learning-models-text-summarization/