July 8th, 2017
(written by lawrence krubner, however indented passages are often quotes). You can contact lawrence at: firstname.lastname@example.org
Evaluating Embeddings: Analogical Reasoning
Embeddings are useful for a wide variety of prediction tasks in NLP. Short of training a full-blown part-of-speech model or named-entity model, one simple way to evaluate embeddings is to directly use them to predict syntactic and semantic relationships like king is to queen as father is to ?. This is called analogical reasoning and the task was introduced by Mikolov and colleagues .
The choice of hyperparameters can strongly influence the accuracy on this task. To achieve state-of-the-art performance on this task requires training over a very large dataset, carefully tuning the hyperparameters and making use of tricks like subsampling the data, which is out of the scope of this tutorial.
Optimizing the Implementation
Our vanilla implementation showcases the flexibility of TensorFlow. For example, changing the training objective is as simple as swapping out the call to tf.nn.nce_loss() for an off-the-shelf alternative such as tf.nn.sampled_softmax_loss(). If you have a new idea for a loss function, you can manually write an expression for the new objective in TensorFlow and let the optimizer compute its derivatives. This flexibility is invaluable in the exploratory phase of machine learning model development, where we are trying out several different ideas and iterating quickly.
Once you have a model structure you’re satisfied with, it may be worth optimizing your implementation to run more efficiently (and cover more data in less time). For example, the naive code we used in this tutorial would suffer compromised speed because we use Python for reading and feeding data items — each of which require very little work on the TensorFlow back-end. If you find your model is seriously bottlenecked on input data, you may want to implement a custom data reader for your problem, as described in New Data Formats. For the case of Skip-Gram modeling, we’ve actually already done this for you as an example in models/tutorials/embedding/word2vec.py.
If your model is no longer I/O bound but you want still more performance, you can take things further by writing your own TensorFlow Ops, as described in Adding a New Op. Again we’ve provided an example of this for the Skip-Gram case models/tutorials/embedding/word2vec_optimized.py. Feel free to benchmark these against each other to measure performance improvements at each stage.
In this tutorial we covered the word2vec model, a computationally efficient model for learning word embeddings. We motivated why embeddings are useful, discussed efficient training techniques and showed how to implement all of this in TensorFlow. Overall, we hope that this has show-cased how TensorFlow affords you the flexibility you need for early experimentation, and the control you later need for bespoke optimized implementation.