Review — context2vec: Learning Generic Context Embedding with Bidirectional LSTM

Using Bidirectional LSTM Instead of Averaging in Word2Vec

**A 2D illustration of context2vec’s embedded space and similarity metrics. Triangles and circles denote sentential context embeddings and target word embeddings, respectively**

In this story, context2vec: Learning Generic Context Embedding with Bidirectional LSTM, (context2vec), by Bar-Ilan University, is briefly reviewed. In this paper:

A bidirectional LSTM is proposed for efficiently learning a generic context embedding function from large corpora.

This is a paper in 2016 CoNLL with over 400 citations. (Sik-Ho Tsang @ Medium)

Outline

CBOW in Word2Vec
Bidirectional LSTM in context2vec
Experimental Results

1. CBOW in Word2Vec

In brief, CBOW represents the context around a target word as a simple average of the embeddings of the context words in a window around it.
The context window can be larger, e.g. extend to [-5, 5] to have 5 previous words and 5 future words.
Obviously, averaging is not good enough, the weighting should be depending on the contexts around the word.

2. Bidirectional LSTM in context2vec

A bidirectional LSTM recurrent neural network, feeding one LSTM network with the sentence words from left to right, and another from right to left.
The parameters of these two networks are completely separate, including two separate sets of left-to-right and right-to-left context word embeddings.
The LSTM output vector representing its left-to-right context (“John”) with the one representing its right-to-left context (“a paper”), are concatenated. With this, the relevant information in the sentential context can be captured:

Next, this concatenated vector is fed into a multi-layer perceptron (MLP) to be capable of representing non-trivial dependencies between the two sides of the context:

where MLP is two-layer MLP:

Negative Sampling is used to optimize the network:

where the summation goes over each word token t in the training corpus and its corresponding (single) sentential context c, and σ is the sigmoid function. t1, …, tk are the negative samples, independently sampled from a smoothed version of the target words unigram distribution:

where 0⩽α<1 is a smoothing factor, which increases the probability of rare words for Negative Sampling.
Some details in the network:

3. Experimental Results

3.1. MSCC Corpus Development Set

**Development set results (iters+ denotes the best model found when running more training iterations with α = 0.75)**

context2vec outperforms AWE, which is another SOTA approach.
Training the proposed models with more iterations and it is found that with 3 iterations over the ukWaC corpus and 10 iterations over the MSCC corpus, some further improvement can be obsereved.

**Test set results (c2v is context2vec)**

As shown above, context2vec substantially outperforms AWE across all benchmarks.
S-1/S-2 stand for the best/second-best prior result reported for the benchmark. context2vec either surpass or almost reach the state-of-the-art on all benchmarks.