Document/Sentence Similarity Solution
summary
A working set of three demos compares actual methods: a book‑title nearest‑neighbor, a toy search engine, and an STS benchmark evaluation. Text converts to vectors with tf-idf, averaged word embeddings, SIF, Universal Sentence Encoder, and sentence-transformers (Hugging Face / SBERT); cosine similarity measures pairs and Pearson correlation scores models. Results show SBERT-style transformers score best, USE second, SIF beats simple averaging, and tf-idf performs worst, with SBERT noticeably slower and costlier; PCA/t‑SNE/UMAP and heatmaps visualize the clusters.
AI-written summary · about
chapters (26)
- 0:00Start
- 1:32Agenda
- 6:09What is STS - Semantic Textual Similarity
- 9:22Demo of Solved Problems
- 10:28Textural similarity measures
- 13:42Compare Similarity Measure
- 16:52Sentence pre-processing using NLP
- 20:53Feature engineering using Sentence Embedding techniques
- 23:08Evolution of Text Embedding Models
- 29:56Word Embedding - Word2Vec
- 31:02Sentence Embedding - by averaging word embeddings
- 32:46Sentence Embedding - by using Smooth Inverse Frequency (SIF)
- 35:48Sentence Embedding example
- 37:35USE architecture
- 39:16S-Bert Encoder
- 41:51Visualizing Sentence embeddings
- 48:56Dimension Reduction library/API details
- 49:27Example of PCA Visualization (tf-idf)
- 50:17Example of t-SNE Visualization (sif)
- 50:46Example of MDS Visualization (use)
- 51:31Example of UMAP Visualization (s-bert)
- 51:52Heatmap of TF-IDF computed similarity
- 57:35Heatmap of SIF Word Embedding computed similarity
- 57:49Heatmap of USE computed similarity
- 58:15Heatmap of S-BERT computed similarity
- 58:34Validation of Similarity Results
related talks
45mApplying the four step "Embed, Encode, Attend, Predict" framework to predict document similarity
38mHow similar are two pieces of text?
41mSemantic Multimodal Multilingual Similarity Engine
45mHybrid search: Greater than the sum of its parts?
42mThe Machine Understands Text — Learning and Search at Micro and Macro Scale
01h 26mBagging to BERT- A tour of applied NLP
01h 01mWord Embeddings - An Alternative & Efficient Approach to Search for Docs
28m"Embeddings: How Computers Learned to Read"