Neural Networks and Deep Learning

Lecture 6 · Lab: predict, test, prove

Five rounds on how text becomes numbers. In each round you first predict what will happen and lock your answer, then test it on the course data, and only then see the proof. The data are those of the lecture: 400 speeches of the European Central Bank (2022–2026), on which we trained our own word2vec and GloVe, and the Financial PhraseBank sentences.

Charts stay empty until you lock a prediction. In round 2, drag the word space to turn it; scroll or pinch to zoom.

Round 1 · A few words do most of the work

  1. 1 Predict
  2. 2 Test
  3. 3 Prove

Round 2 · Who are the neighbours?

  1. 1 Predict
  2. 2 Test
  3. 3 Prove

Round 3 · Arithmetic with words

  1. 1 Predict
  2. 2 Test
  3. 3 Prove

Round 4 · Averaging embeddings or counting words?

  1. 1 Predict
  2. 2 Test
  3. 3 Prove

Round 5 · When the annotators disagree

  1. 1 Predict
  2. 2 Test
  3. 3 Prove

ECB speeches: 400 speeches, 8 January 2022 to 30 October 2026, 1,289,896 tokens. word2vec: skip-gram with negative sampling, 100 dimensions, window 5, words seen at least 5 times; GloVe on the same text; both trained in PyTorch. The 3D view is a principal component projection of the 32 words of four topic groups. Financial PhraseBank (Malo et al., 2014): 4,838 sentences, 5 random splits; finance tweets: 3 seeds. Code: lectures/L6/analysis_L6.py and build_interactive_L6.py on GitHub.