Understanding Names with Neural Networks - Session 2

Explore the challenges and innovative solutions in cross-language name matching during this 59-minute webinar on neural network applications. Delve into the limitations of traditional methods like edit distance and Hidden Markov Models, and discover how deep neural networks significantly improve accuracy in English/Japanese name matching. Learn about sequence-to-sequence models, Long Short-Term Memory cells, and Convolutional Neural Networks as applied to transliteration and name scoring. Gain insights into BasisTech's Rosette capabilities and the practical implications of these advanced techniques for consumer and governmental domains.

Syllabus

Intro
ABOUT BASIS TECHNOLOGY
Rosette Capabilities
Names are a Challenge
Task: Name Matching
Name Matching Algorithms
Step One: Modeling Sequences of Characters
Step Two: Modeling Transliterations
Issues with HMM-Based Name Matching
What does an HMM Actually Do?
How Would You Transliterate a Name?
The Antidote: Sequence-to-Sequence (seq2seq)
Neural Network of Choice: Long Short-Term Memory (LSTM) Cells
Step One: Learning to Transliterate with seq2seq
Step Two: Running the Transliterator in Reverse to Score
How Can We Produce a Score?
Processing Time on Name Pairs (seconds)
Faster seq2seq with a Convolutional Neural Network CNNO
Convolutional Neural Net (CNN)
CNN in Natural Language Processing
What Does This Tell Us?