README.md

Sentence Similarity Calculator

This repo contains various ways to calculate the similarity between source and target sentences. You can choose the pre-trained models you want to use such as ELMo, BERT and Universal Sentence Encoder (USE).

And you can also choose the method to be used to get the similarity:

1. Cosine similarity
2. Manhattan distance
3. Euclidean distance
4. Angular distance
5. Inner product
6. TS-SS score
7. Pairwise-cosine similarity
8. Pairwise-cosine similarity + IDF

You can experiment with (The number of models) x (The number of methods) combinations!

Installation

• This project is developed under conda enviroment
• After cloning this repository, you can simply install all the dependent libraries described in requirements.txt with bash install.sh

conda create -n sensim python=3.7
conda activate sensim
git clone https://github.com/Huffon/sentence-similarity.git
cd sentence-similarity
bash install.sh

Usage

• To test your own sentences, you should fill out corpus.txt with sentences as below:

I ate an apple.
I went to the Apple.
I ate an orange.
...

• Then, choose the model and method to be used to calculate the similarity between source and target sentences

python sensim.py
    --model    MODEL_NAME  [use, bert, elmo]
    --method   METHOD_NAME [cosine, manhattan, euclidean, inner,
                            ts-ss, angular, pairwise, pairwise-idf]
    --verbose  LOG_OPTION (bool)

Examples

• In this section, you can see the example result of sentence-similarity
• As you know, there is a no silver-bullet which can calculate perfect similarity between sentences
• You should conduct various experiments with your dataset
  • Caution: TS-SS score might not fit with sentence similarity task, since this method originally devised to calculate the similarity between long documents
• Result:

References

Papers

• Universal Sentence Encoder
• Deep contextualized word representations
• BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding
• Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks
• BERTScore: Evaluating Text Generation with BERT
• A Hybrid Geometric Approach for Measuring Similarity Level Among Documents and Document Clustering

Libraries

• TF-hub's Universal Sentence Encoder
• Allen NLP's ELMo
• Sentence Transformers
• BERTScore
• Vector Similarity

Articles

• An Overview of Sentence Embedding Methods
• Comparing Sentence Similarity Methods
• The Current Best of Universal Word Embeddings and Sentence Embeddings