• Encoding sentences into embedding vectors for transfer learning to other NLP tasks

• Transfer learning using sentence embeddings tends to outperform word level transfer

Two variants: transformer encoder and deep averaging network

Transformer Encoder

• targets high accuracy at the cost of greater model complexity
• input is lowercased Penn Treebank tokenized string (see J&M chap. 2.4)
• compute context aware representations of words
• compute element-wise sum at each word position to get fixed length sentence encoding vector
• divide by square root of sentence length to mitigate length effects
• time complexity is $O(n^2)$ / space complexity is $O(n^2)$

Deep Averaging Network (DAN)

(Iyyer et al., 2015)

• targets efficient inference with slightly reduced accuracy
• input embeddings for words and bi-grams are first averaged together
• and then passed through a feedforward deep neural network to produce sentence embeddings
• time complexity is $O(n)$ / space complexity is $O(1)$
• space complexity is dominated by the parameters used to store unigram and bigram embeddings (for short sentences, can be twice the memory usage of transformer model)

Multi-task learning

single encoding model used to feed multiple downstream tasks

• skip-thought task (unsupervised from running text)
• conversational input-response task
• classification tasks (supervised)

Transfer Learning

• For sentence classification transfer tasks, the output of the transformer and DAN sentence encoders are provided to a task specific deep neural net.
• For pairwise semantic similarity task, compute cosine similarity of the sentence embeddings produced by both encoder variants and then apply arccos to convert into angular distance (performs better than raw cosine similarity) $\text{sim}(u,v) = 1-\arccos \big( \frac{u\cdot v}{\lVert u\rVert \lVert v \rVert} \big)/\pi$

Results

• hyperparemeters tuned using Google Vizier
• assess bias in encoding models by evaluating strength of associations learned on WEAT word lists (Word Embeddings Association Test)
• sentence-level + word-level (word2vec skip-gram embeddings) transfer > sentence-level transfer only > word-level transfer only
• transfer learning is most critical when training data for a target task is limited
• as training set size increases, marginal improvement of using transfer learning decreases

Note: switching to Sentence Piece vocabulary instead of words significantly reduces vocabulary size, which is a major contributor of model sizes (good for on-device or browser-based implementations)