Natural language modelling- BlazingText algorithm
summary
Demonstrates producing word embeddings from a Tamil Wikipedia dump using the BlazingText algorithm packaged in Amazon SageMaker, with a distributed training run that completes in about 32.86 seconds on four c4.2xlarge instances. Claims BlazingText scales to billions of words and up to ~50 million words/sec in distributed mode, and yields sensible neighbors (music/song distance ~6.17, politics/leader ~5.8). Model is exported to S3 and deployable with one line of code; t‑SNE visualization confirms coherent clusters.
AI-written summary · about
chapters (9)
- 0:00Natural language modeling intro
- 2:35Word embeddings necessity
- 7:10Word2Vec algorithm
- 11:27Amazon SageMaker overview
- 23:52BlazingText algorithm intro
- 29:33Demo setup and data prep
- 33:54Training with BlazingText
- 40:20Model deployment demo
- 45:23Embedding visualization
related talks
31mTaking AI/ML to the next level: Text!
25mGenerative Text Modelling: Scratching the surface
32mGenerative AI Text and Multimodal Embedding Models for Real World Use Cases
22mBoring but Effective: WebAI That Ships
45mDeep Learning and NLP with Spark
34mModelling a text corpus using Deep Boltzmann Machines
45mDeep Learning for Natural Language Processing
57mMachine Learning with JavaScript