This is a work-in-progress, but much-improved, re-implementation of the linguistically-informed self-attention (LISA) model described in the following paper:
Emma Strubell, Patrick Verga, Daniel Andor, David Weiss, and Andrew McCallum. Linguistically-Informed Self-Attention for Semantic Role Labeling. Conference on Empirical Methods in Natural Language Processing (EMNLP). Brussels, Belgium. October 2018.
To exactly replicate the results in the paper at the cost of an unpleasantly hacky codebase, you can use the original LISA code here.
- >= Python 3.6
- >= TensorFlow 1.9 (tested up to 1.12)
- Get pre-trained word embeddings (GloVe):
wget -P embeddings http://nlp.stanford.edu/data/glove.6B.zip unzip -j embeddings/glove.6B.zip glove.6B.100d.txt -d embeddings - Get CoNLL-2005 data in the right format using this repo. Follow the instructions all the way through further preprocessing.
- Make sure the correct data paths are set in
config/conll05.conf
To train a model with save directory model using the configuration conll05-lisa.conf:
bin/train.sh config/conll05-lisa.conf --save_dir model
To evaluate the latest checkpoint saved in the directory model:
bin/evaluate.sh config/conll05-lisa.conf --save_dir model
To evaluate the best1 checkpoint so far, saved in the directory model (with id 1554216594):
bin/evaluate-exported.sh config/conll05-lisa.conf --save_dir model/export/best_exporter/1554216594
The bin/train.sh script calls src/train.py with parameters specified in top-level configs (i.e. conll05-lisa.conf) which is the entry point for training. The following table describes the command line parameters that may be passed to src/train.py to configure training:
| Name | Type | Description | Default value |
|---|---|---|---|
train-files |
string | Comma-separated list of training data files. | None |
dev-files |
string | Comma-separated list of development data files. | None |
save-dir |
string | Directory to save models, outputs, etc. If the directory already exists and contains a trained model, training will restart where it left off. Vocabularies will be re-used. | None |
transition_stats |
string | File containing pre-computed transition statistics between labels. Tab-separated file with one label-label-probability triple per line. | None |
hparams |
string | Comma separated list of name=value hyperparameter settings. |
None |
debug |
string | Whether to run in debug mode: a little faster and smaller. | False |
data_config |
string | Path to data configuration json. | None |
model_configs |
string | Comma-separated list of paths to model configuration json. | None |
task_configs |
string | Comma-separated list of paths to data configuration json. | None |
layer_configs |
string | Comma-separated list of paths to data configuration json. | None |
attention_configs |
string | Comma-separated list of paths to attention configuration json. | None |
keep_k_best_models |
int | Number of best models to keep. | 1 |
best_eval_key |
string | Key corresponding to the evaluation to be used for determining early stopping. The value must correspond to a named eval under the eval_fns entry in a task config. |
None |
The following table lists optimization/training hyperparameters that can be set through the hparams command line flag. Hyperparameters are initialized to the default values are defined in src/constants.py. Then, these are overridden by hyperparameters set in the model config (e.g., glove_basic.json). Finally, these are overridden by hyperparameters specified at the command line. Hyperparameter loading is implemented in src/train_utils.py.
| Name | Type | Description | Default value |
|---|---|---|---|
learning_rate |
float | Initial learning rate. | 0.04 |
beta1 |
float | Adam first moment decay rate. | 0.9 |
beta2 |
float | Adam second moment decay rate. | 0.98 |
epsilon |
float | Adam epsilon. | 1e-12 |
decay_rate |
float | Exponential rate of decay for learning rate. | 1.5 |
use_nesterov |
boolean | Whether to use Nesterov momentum in Adam. | true |
decay_steps |
int | If warmup_steps is not set, perform stepwise decay of learning rate every this many steps. |
5000 |
warmup_steps |
int | Number of training steps to linearly increase learning rate before exponential decay. | 8000 |
batch size |
int | Approximate number of sentences per batch. | 256 |
shuffle_buffer_multiplier |
int | Value to multiply by batch size to determine buffer size for efficient shuffling of examples during training. Higher means better shuffles, lower means less initial time required to fill shuffle buffer. | 100 |
eval_throttle_secs |
int | Do not run evaluation unless at least this many seconds have passed since the last evaluation. | 1000 |
eval_every_steps |
int | Evaluate every this many steps. | 1000 |
num_train_epochs |
int | Iterate through the full training data this many times. | 10000 |
gradient_clip_norm |
float | Clip gradients to this maximum value. | 5.0 |
label_smoothing |
float | Amount of label corruption for smoothing. Smoothing not performed if this value is 0. | 0.1 |
moving_average_decay |
float | Rate of decay for moving average of model parameters. Averaging not performed if this value is 0. | 0.999 |
average_norms |
boolean | Whether to average variables representing norms in parameter averaging. | false |
input_dropout |
float | Dropout rate on input layer (embeddings). | 1.0 |
bilinear_dropout |
float | Dropout rate used in bilinear classifier. | 1.0 |
mlp_dropout |
float | Dropout used in MLP layers | 1.0 |
attn_dropout |
float | Dropout rate on attention in transformer. | 1.0 |
ff_dropout |
float | Dropout rate in feed-forward layer in transformer. | 1.0 |
prepost_dropout |
float | Dropout rate applied before and after the feed-forward part of transformer layer. | 1.0 |
random_seed |
int | Random seed to use for training. | time.time() |
Model hyperparameters (e.g. layer size, number of self-attention heads) are set in the model config json.
TODO
LISA model configuration is defined through a combination of configuration files. A top-level config defines a specific model configuration and dataset by setting other configurations. Top-level configs are written in bash, and bottom-level configs are written in json. Here is an example top-level config, conll05-lisa.conf, which defines the basic LISA model and CoNLL-2005 data:
# use CoNLL-2005 data
source config/conll05.conf
# take glove embeddings as input
model_configs=config/model_configs/glove_basic.json
# joint pos/predicate layer, parse heads and labels, and srl
task_configs="config/task_configs/joint_pos_predicate.json,config/task_configs/parse_heads.json,config/task_configs/parse_labels.json,config/task_configs/srl.json"
# use parse in attention
attention_configs="config/attention_configs/parse_attention.json"
# specify the layers
layer_configs="config/layer_configs/lisa_layers.json"
And the top-level data config for the CoNLL-2005 dataset that it loads, conll05.conf:
data_config=config/data_configs/conll05.json
data_dir=$DATA_DIR/conll05st-release-new
train_files=$data_dir/train-set.gz.parse.sdeps.combined.bio
dev_files=$data_dir/dev-set.gz.parse.sdeps.combined.bio
test_files=$data_dir/test.wsj.gz.parse.sdeps.combined.bio,$data_dir/test.brown.gz.parse.sdeps.combined.bio
Note that $DATA_DIR is a bash global variable, but all the other variables are defined in these configs.
There are five types of bottom-level configurations, specifying different aspects of the model:
- data configs: Data configs define a mapping from columns in a one-word-per-line formatted file (e.g. the CoNLL-X format) to named features and labels that will be provided to the model as batches.
- model configs: Model configs define hyperparameters, both model hyperparameters, like various embedding dimensions, and optimization hyperparameters, like learning rate. Optimization hyperparameters can be reset at the command line using the
hparamscommand line parameter, which takes a comma-separated list ofname=valuehyperparameter settings. Model hyperparameters cannot be redefined in this way, since this would invalidate a serialized model. - task configs: Task configs define a task: label, evaluation, and how predictions are formed from the model. Each task (e.g. SRL, parse edges, parse labels) should have its own task config.
- layer configs: Layer configs attach tasks to layers, defining which layer representations should be trained to predict named labels (from the data config). The number of layers in the model is determined by the maximum depth listed in layer configs.
- attention configs (optional): Attention configs define special attention functions which replace attention heads, i.e. syntactically-informed self attention. Omitting any attention configs results in a model performing simple single- or multi-task learning.
How these different configuration files work is specified in more detail below.
An full example data config can be seen here: conll05.json.
Each top-level entry in the json defines a named feature or label that will be provided to the model. The following table describes the possible parameters for configuring how each input is interpreted.
| Field | Type | Description | Default value |
|---|---|---|---|
conll_idx |
int or list | Column in the data file corresponding to this input. | N/A (required) |
vocab |
string | Name of the vocabulary used to map this (string) input to int. | None (output of converter is int) |
type |
string | Type of conll_idx. Possible types are: range, other (int/list). "range" can be used to specify that a variable-length range of columns should be read in at once and passed to the converter. Otherwise, the given single int or list of columns is read in and passed to the converter. |
"other" (int/list) |
feature |
boolean | Whether this input should be used as a feature, i.e. provided to the model as input. | false |
label |
boolean | Whether this input should be used as a label, i.e. provided to the model as a label. | false |
updatable |
boolean | Whether this vocab should be updated after its initial creation (i.e. after creating a vocab based on the training data). | false |
converter |
json | A json object defining a function (name and, optionally, parameters) for converting the raw input. These functions are defined in src/data_converters.py. |
idx_list_converter |
oov |
boolean | Whether an OOV entry should be added to this input's vocabulary. |
false |
The data config specifies a converter function and vocabulary for each desired column in the input data file. For each entry in the data config and each line in the input file, the column values specified by conll_idx are read in and provided to the given converter. Data generators, which take the data config and data file as input to perform this mapping, are defined in src/data_generator.py.
New converter functions can be defined in src/data_converters.py. At a minimum, every converter function takes two parameters: split_line, the current line in the data file split by whitespace, and idx, the value of conll_idx. Converters may also take additional parameters, whose values are defined via the params field in the converter json object. The output of a converter is a list of strings.
For example, the default converter, idx_list_converter, simply takes a single column index or list of indices and returns a list containing the corresponding column values in the input file:
def idx_list_converter(split_line, idx):
if isinstance(idx, int):
return [split_line[idx]]
return [split_line[i] for i in idx]When a vocab is specified for an entry in the data config, that vocab is used to map the string output of the converter to integer values suitable for features/labels in a TensorFlow model.2 This mapping occurs in the map_strings_to_ints function in src/dataset.py.
- TODO: vocab initialization
- TODO: pre-trained word embeddings
TODO
TODO
TODO
TODO
1: "Best" is determined by best_eval_key, with default value for a given dataset in the top-level data config, e.g. config/conll05.conf. The value of best_eval_key must correspond to a named eval under the eval_fns entry in a task config. ↩︎
2: If no vocab is specified, then it's assumed that the output of the converter can be interpreted as an integer. ↩︎