Explanations, formulas, visualisations:
- Lena Voita's blog: Text Classification
- Goldberg 9, 10, 11
- Eisenstein 3.4, 6
- Jurafsky-Martin 9
- RNNs are designed to process sequences, when the order of symbols is important
- Hidden layers of RNNs model the whole sequence, each token in the input is represented with a hidden state
- The hidden states can be seen as layers of a deep network (often called units) connected via the RNN abstraction
- The RNN abstraction: current state = current input + previous state
- Gating is invented to control how much history is allowed to pass to the next state, gated RNN units have more structure than simple RNN unit
- by design uni-directional, in NLP mostly used as bi-directional: concatenation of the left-to-right and right-to-left pass
- can output a symbol at each step or at the end of the sequence
- LSTM is a kind of gating: each state consists of a hidden and a memory (also called context) component
- Hidden component is the main unit, memory is additional unit that helps gating
- Gates: input, forget, output
- designed to identify relevant input features
- specifically in NLP: relevant segments of text are found with 1D convolutions
- convolution produces one vector representation for each n-gram
- each filter is responsible for one cell in the vector, dimensionality of n-gram vector is the number of filters
- n-gram representations are then pooled into a single vector representing the whole text
- this vector is input to a classifier
- n-gram size ≠ filter size!
- Typically CNNs are better suited for the sentence-level tasks (e.g. spam filtering, sentiment), while LSTMs for token level tasks (e.g. NER, POS tagging)
- LSTMs are also often used for sentence level tasks: the last state is a representation of the whole sequence
- The use of CNNs for token-level tasks is possible but not common