Siamese LSTM in Keras: Learning Character-Based Phrase
summary
Claims a working pipeline that maps noisy job-title strings into a 5,000+ class taxonomy by learning character embeddings with a B-LSTM encoder trained as a Siamese network in Keras 1.1.1. Training uses synthetic augmentation, mined synonyms, and a crucial neighbor sampling strategy to surface hard negatives. Results show measurable gains over trigram/string baselines, better handling of typos, abbreviations and synonyms, and concrete evaluation via unit-style tests and labeled prediction sets.
AI-written summary · about
chapters (11)
- 0:00Introduction
- 2:34Job title taxonomy mapping
- 5:30Character Bi-LSTM encoder
- 8:50Training with cosine distance
- 11:58Keras implementation
- 17:13Training challenges
- 19:46Data augmentation
- 22:18Neighborhood sampling
- 24:28Model evaluation
- 27:20Results and improvements
- 28:48Conclusion
related talks
40mVisual concept learning from few images
39mDeep Learning vs. Conventional Machine Learning
01h 26mFrom RNNs to Transformers: A Journey through Language Modeling
54mNatural Language Processing with Deep Learning and TensorFlow
01h 30mIntroduction to Deep Learning & Natural Language Processing
34mClassifying Search Queries Without User Click Data
16mMachine Learning on Source Code
06mWeakly Supervised Multi-Task Representation Learning for Human Activity Analysis Using Wearables