DependencyTreeRnn

Dependency tree-based RNN

Copyright (c) 2014-2015 Piotr Mirowski, Andreas Vlachos

Please refer to the following paper: Piotr Mirowski, Andreas Vlachos "Dependency Recurrent Neural Language Models for Sentence Completion" ACL 2015

Installation

1. Modify the path to the BLAS header (cblas.h) file, i.e., $BLASINCLUDE and the BLAS path, i.e., $BLASFLAGS, in file Makefile. Alternatively, make your own version of that Makefile.
2. Build the project:

> make

or, using your custom Makefile:

> make -f YOUR_OWN_MAKEFILE

Note that the .o objects are stored in directory build/ and the executable is ./RnnDependencyTree

Sample training script

Shell script train_rnn_holmes_debug.sh trains an RNN on a subset of a few books. You need to modify the path to where the JSON book files are stored.

Important hyperparameters

1. Parameters relative to the dataset:

• train (string) Training data file (pure text)
• valid (string) Validation data file (pure text), using during training
• test (string) Test data file (pure text)
• sentence-labels (string) Validation/test sentence labels file (pure text)
• path-json-books (string) Path to the book JSON files
• min-word-occurrence (int) Mininum word occurrence to include word into vocabulary [default: 5]
• independent (bool) Is each line in the training/testing file independent? [default: true]

2. Parameters relative to the dependency labels

• feature-labels-type (int) Dependency parsing labels:
  • 0 = none, use words only
  • 1 = concatenate label to word
  • 2 = use features in the feature vector, separate from words
• feature-gamma (double) Decay weight for features consisting of label vectors [default: 0.9].
  • Values up to about 1.3 can be accepted (beyond that, the perplexity seems to become very large).
  • f(t) is a vector with D elements (e.g., D=44 types of dependency labels)
  • f(t) <- gamma * f(t-1), then set element at current label to 1
  • This value could be important for changing the weight given to dependency labels.
    • A value larger than 1 means that labels further past in time count more than those immediately in the past.
    • 1 means that there is no decay.
    • A value between 0 and 1 means that there is some decay.
    • 0 means that the decay is immediate.

3. RNN architecture parameters

• rnnlm (string) RNN language model file to use (save in training / read in test)
• classes (int) Number of word classes used in hierarchical softmax [default: 200].
  • If vocabulary size if W, choose C around sqrt(W).
  • C = W means 1 class per word.
  • C = 1 means standard softmax.
• hidden (int) Number of nodes in the hidden layer [default: 100].
  • Try to go higher, perhaps up to 1000 (for 1M-word vocabulary).
  • Linear impact on speed.
• direct (int) Size of max-entropy hash table storing direct n-gram connections, in millions of entries [default: 0].
  • Basically, direct=1000 means that 1000x10000000 = 1G direct connections between context words and target word are considered.
  • However, it is not a proper hashtable (which would take too much memory) but a simple vector of 1G entries, with a hashing function that hashed into specific entries in that vector. Hash collisions are totally ignored.
  • Try using direct=1000 or even 2000 hashes if possible.
• direct-order (int) Order of direct n-gram connections; 2 is like bigram max entropy features [default: 3].
  • It works on tokens only, and values of 4 or beyond did not bring improvement in others LM tasks.
• compression (int) Number of nodes in the compression layer between the hidden and output layers [default: 0]

4. Training parameters

• alpha (double) Initial learning rate during gradient descent [default: 0.1]
• beta (double) L-2 norm regularization coefficient during gradient descent [default: 0.0000001]
• min-improvement (double) Minimum improvement before learning rate decreases [default: 1.001]
• bptt (int) Number of steps to propagate error back in time [default: 4]
• bptt-block (int) Number of time steps after which the error is backpropagated through time [default: 10]
• gradient-cutoff (double) Value beyond whih the gradients are clipped, used to avoid exploding gradients [default: 15]

5. Additional parameters

• debug (bool) Debugging level [default: false]