Find file History
Latest commit bc70271 Feb 24, 2017 @bogatyy bogatyy committed with calberti Fix Bazel version check (#1069)

SyntaxNet: Neural Models of Syntax.

A TensorFlow implementation of the models described in Andor et al. (2016).

Update: Parsey models are now available for 40 languages trained on Universal Dependencies datasets, with support for text segmentation and morphological analysis.

At Google, we spend a lot of time thinking about how computer systems can read and understand human language in order to process it in intelligent ways. We are excited to share the fruits of our research with the broader community by releasing SyntaxNet, an open-source neural network framework for TensorFlow that provides a foundation for Natural Language Understanding (NLU) systems. Our release includes all the code needed to train new SyntaxNet models on your own data, as well as Parsey McParseface, an English parser that we have trained for you, and that you can use to analyze English text.

So, how accurate is Parsey McParseface? For this release, we tried to balance a model that runs fast enough to be useful on a single machine (e.g. ~600 words/second on a modern desktop) and that is also the most accurate parser available. Here's how Parsey McParseface compares to the academic literature on several different English domains: (all numbers are % correct head assignments in the tree, or unlabelled attachment score)

Model News Web Questions
Martins et al. (2013) 93.10 88.23 94.21
Zhang and McDonald (2014) 93.32 88.65 93.37
Weiss et al. (2015) 93.91 89.29 94.17
Andor et al. (2016)* 94.44 90.17 95.40
Parsey McParseface 94.15 89.08 94.77

We see that Parsey McParseface is state-of-the-art; more importantly, with SyntaxNet you can train larger networks with more hidden units and bigger beam sizes if you want to push the accuracy even further: Andor et al. (2016)* is simply a SyntaxNet model with a larger beam and network. For futher information on the datasets, see that paper under the section "Treebank Union".

Parsey McParseface is also state-of-the-art for part-of-speech (POS) tagging (numbers below are per-token accuracy):

Model News Web Questions
Ling et al. (2015) 97.44 94.03 96.18
Andor et al. (2016)* 97.77 94.80 96.86
Parsey McParseface 97.52 94.24 96.45

The first part of this tutorial describes how to install the necessary tools and use the already trained models provided in this release. In the second part of the tutorial we provide more background about the models, as well as instructions for training models on other datasets.



Running and training SyntaxNet models requires building this package from source. You'll need to install:

  • python 2.7:
    • python 3 support is not available yet
  • bazel:
    • version 0.4.3
    • follow the instructions here
    • Alternately, Download bazel (0.4.3) <.deb> from for your system configuration.
    • Install it using the command: sudo dpkg -i <.deb file>
    • Check for the bazel version by typing: bazel version
  • swig:
    • apt-get install swig on Ubuntu
    • brew install swig on OSX
  • protocol buffers, with a version supported by TensorFlow:
    • check your protobuf version with pip freeze | grep protobuf
    • upgrade to a supported version with pip install -U protobuf==3.0.0b2
  • mock, the testing package:
    • pip install mock
  • asciitree, to draw parse trees on the console for the demo:
    • pip install asciitree
  • numpy, package for scientific computing:
    • pip install numpy

Once you completed the above steps, you can build and test SyntaxNet with the following commands:

  git clone --recursive
  cd models/syntaxnet/tensorflow
  cd ..
  bazel test syntaxnet/... util/utf8/...
  # On Mac, run the following:
  bazel test --linkopt=-headerpad_max_install_names \
    syntaxnet/... util/utf8/...

Bazel should complete reporting all tests passed.

You can also compile SyntaxNet in a Docker container using this Dockerfile.

To build SyntaxNet with GPU support please refer to the instructions in issues/248.

Note: If you are running Docker on OSX, make sure that you have enough memory allocated for your Docker VM.

Getting Started

Once you have successfully built SyntaxNet, you can start parsing text right away with Parsey McParseface, located under syntaxnet/models. The easiest thing is to use or modify the included script syntaxnet/, which shows a basic setup to parse English taking plain text as input.

Parsing from Standard Input

Simply pass one sentence per line of text into the script at syntaxnet/ The script will break the text into words, run the POS tagger, run the parser, and then generate an ASCII version of the parse tree:

echo 'Bob brought the pizza to Alice.' | syntaxnet/

Input: Bob brought the pizza to Alice .
brought VBD ROOT
 +-- Bob NNP nsubj
 +-- pizza NN dobj
 |   +-- the DT det
 +-- to IN prep
 |   +-- Alice NNP pobj
 +-- . . punct

The ASCII tree shows the text organized as in the parse, not left-to-right as visualized in our tutorial graphs. In this example, we see that the verb "brought" is the root of the sentence, with the subject "Bob", the object "pizza", and the prepositional phrase "to Alice".

If you want to feed in tokenized, CONLL-formatted text, you can run --conll.

Annotating a Corpus

To change the pipeline to read and write to specific files (as opposed to piping through stdin and stdout), we have to modify the to point to the files we want. The SyntaxNet models are configured via a combination of run-time flags (which are easy to change) and a text format TaskSpec protocol buffer. The spec file used in the demo is in syntaxnet/models/parsey_mcparseface/context.pbtxt.

To use corpora instead of stdin/stdout, we have to:

  1. Create or modify an input field inside the TaskSpec, with the file_pattern specifying the location we want. If the input corpus is in CONLL format, make sure to put record_format: 'conll-sentence'.
  2. Change the --input and/or --output flag to use the name of the resource as the output, instead of stdin and stdout.

E.g., if we wanted to POS tag the CONLL corpus ./wsj.conll, we would create two entries, one for the input and one for the output:

input {
  name: 'wsj-data'
  record_format: 'conll-sentence'
  Part {
    file_pattern: './wsj.conll'
input {
  name: 'wsj-data-tagged'
  record_format: 'conll-sentence'
  Part {
    file_pattern: './wsj-tagged.conll'

Then we can use --input=wsj-data --output=wsj-data-tagged on the command line to specify reading and writing to these files.

Configuring the Python Scripts

As mentioned above, the python scripts are configured in two ways:

  1. Run-time flags are used to point to the TaskSpec file, switch between inputs for reading and writing, and set various run-time model parameters. At training time, these flags are used to set the learning rate, hidden layer sizes, and other key parameters.
  2. The TaskSpec proto stores configuration about the transition system, the features, and a set of named static resources required by the parser. It is specified via the --task_context flag. A few key notes to remember:

    • The Parameter settings in the TaskSpec have a prefix: either brain_pos (they apply to the tagger) or brain_parser (they apply to the parser). The --prefix run-time flag switches between reading from the two configurations.
    • The resources will be created and/or modified during multiple stages of training. As described above, the resources can also be used at evaluation time to read or write to specific files. These resources are also separate from the model parameters, which are saved separately via calls to TensorFlow ops, and loaded via the --model_path flag.
    • Because the TaskSpec contains file path, remember that copying around this file is not enough to relocate a trained model: you need to move and update all the paths as well.

Note that some run-time flags need to be consistent between training and testing (e.g. the number of hidden units).

Next Steps

There are many ways to extend this framework, e.g. adding new features, changing the model structure, training on other languages, etc. We suggest reading the detailed tutorial below to get a handle on the rest of the framework.

Detailed Tutorial: Building an NLP Pipeline with SyntaxNet

In this tutorial, we'll go over how to train new models, and explain in a bit more technical detail the NLP side of the models. Our goal here is to explain the NLP pipeline produced by this package.

Obtaining Data

The included English parser, Parsey McParseface, was trained on the the standard corpora of the Penn Treebank and OntoNotes, as well as the English Web Treebank, but these are unfortunately not freely available.

However, the Universal Dependencies project provides freely available treebank data in a number of languages. SyntaxNet can be trained and evaluated on any of these corpora.

Part-of-Speech Tagging

Consider the following sentence, which exhibits several ambiguities that affect its interpretation:

I saw the man with glasses.

This sentence is composed of words: strings of characters that are segmented into groups (e.g. "I", "saw", etc.) Each word in the sentence has a grammatical function that can be useful for understanding the meaning of language. For example, "saw" in this example is a past tense of the verb "to see". But any given word might have different meanings in different contexts: "saw" could just as well be a noun (e.g., a saw used for cutting) or a present tense verb (using a saw to cut something).

A logical first step in understanding language is figuring out these roles for each word in the sentence. This process is called Part-of-Speech (POS) Tagging. The roles are called POS tags. Although a given word might have multiple possible tags depending on the context, given any one interpretation of a sentence each word will generally only have one tag.

One interesting challenge of POS tagging is that the problem of defining a vocabulary of POS tags for a given language is quite involved. While the concept of nouns and verbs is pretty common, it has been traditionally difficult to agree on a standard set of roles across all languages. The Universal Dependencies project aims to solve this problem.

Training the SyntaxNet POS Tagger

In general, determining the correct POS tag requires understanding the entire sentence and the context in which it is uttered. In practice, we can do very well just by considering a small window of words around the word of interest. For example, words that follow the word ‘the’ tend to be adjectives or nouns, rather than verbs.

To predict POS tags, we use a simple setup. We process the sentences left-to-right. For any given word, we extract features of that word and a window around it, and use these as inputs to a feed-forward neural network classifier, which predicts a probability distribution over POS tags. Because we make decisions in left-to-right order, we also use prior decisions as features in subsequent ones (e.g. "the previous predicted tag was a noun.").

All the models in this package use a flexible markup language to define features. For example, the features in the POS tagger are found in the brain_pos_features parameter in the TaskSpec, and look like this (modulo spacing):

stack(3).word stack(2).word stack(1).word stack.word input.word input(1).word input(2).word input(3).word;
input.digit input.hyphen;
stack.suffix(length=2) input.suffix(length=2) input(1).suffix(length=2);
stack.prefix(length=2) input.prefix(length=2) input(1).prefix(length=2)

Note that stack here means "words we have already tagged." Thus, this feature spec uses three types of features: words, suffixes, and prefixes. The features are grouped into blocks that share an embedding matrix, concatenated together, and fed into a chain of hidden layers. This structure is based upon the model proposed by Chen and Manning (2014).

We show this layout in the schematic below: the state of the system (a stack and a buffer, visualized below for both the POS and the dependency parsing task) is used to extract sparse features, which are fed into the network in groups. We show only a small subset of the features to simplify the presentation in the schematic:


In the configuration above, each block gets its own embedding matrix and the blocks in the configuration above are delineated with a semi-colon. The dimensions of each block are controlled in the brain_pos_embedding_dims parameter. Important note: unlike many simple NLP models, this is not a bag of words model. Remember that although certain features share embedding matrices, the above features will be concatenated, so the interpretation of input.word will be quite different from input(1).word. This also means that adding features increases the dimension of the concat layer of the model as well as the number of parameters for the first hidden layer.

To train the model, first edit syntaxnet/context.pbtxt so that the inputs training-corpus, tuning-corpus, and dev-corpus point to the location of your training data. You can then train a part-of-speech tagger with:

bazel-bin/syntaxnet/parser_trainer \
  --task_context=syntaxnet/context.pbtxt \
  --arg_prefix=brain_pos \  # read from POS configuration
  --compute_lexicon \       # required for first stage of pipeline
  --graph_builder=greedy \  # no beam search
  --training_corpus=training-corpus \  # names of training/tuning set
  --tuning_corpus=tuning-corpus \
  --output_path=models \  # where to save new resources
  --batch_size=32 \       # Hyper-parameters
  --decay_steps=3600 \
  --hidden_layer_sizes=128 \
  --learning_rate=0.08 \
  --momentum=0.9 \
  --seed=0 \
  --params=128-0.08-3600-0.9-0  # name for these parameters

This will read in the data, construct a lexicon, build a tensorflow graph for the model with the specific hyperparameters, and train the model. Every so often the model will be evaluated on the tuning set, and only the checkpoint with the highest accuracy on this set will be saved. Note that you should never use a corpus you intend to test your model on as your tuning set, as you will inflate your test set results.

For best results, you should repeat this command with at least 3 different seeds, and possibly with a few different values for --learning_rate and --decay_steps. Good values for --learning_rate are usually close to 0.1, and you usually want --decay_steps to correspond to about one tenth of your corpus. The --params flag is only a human readable identifier for the model being trained, used to construct the full output path, so that you don't need to worry about clobbering old models by accident.

The --arg_prefix flag controls which parameters should be read from the task context file context.pbtxt. In this case arg_prefix is set to brain_pos, so the paramters being used in this training run are brain_pos_transition_system, brain_pos_embedding_dims, brain_pos_features and, brain_pos_embedding_names. To train the dependency parser later arg_prefix will be set to brain_parser.

Preprocessing with the Tagger

Now that we have a trained POS tagging model, we want to use the output of this model as features in the parser. Thus the next step is to run the trained model over our training, tuning, and dev (evaluation) sets. We can use the` script for this.

For example, the model 128-0.08-3600-0.9-0 trained above can be run over the training, tuning, and dev sets with the following command:

for SET in training tuning dev; do
  bazel-bin/syntaxnet/parser_eval \
    --task_context=models/brain_pos/greedy/$PARAMS/context \
    --hidden_layer_sizes=128 \
    --input=$SET-corpus \
    --output=tagged-$SET-corpus \
    --arg_prefix=brain_pos \
    --graph_builder=greedy \

Important note: This command only works because we have created entries for you in context.pbtxt that correspond to tagged-training-corpus, tagged-dev-corpus, and tagged-tuning-corpus. From these default settings, the above will write tagged versions of the training, tuning, and dev set to the directory models/brain_pos/greedy/$PARAMS/. This location is chosen because the input entries do not have file_pattern set: instead, they have creator: brain_pos/greedy, which means that will construct new files when called with --arg_prefix=brain_pos --graph_builder=greedy using the --model_path flag to determine the location.

For convenience, also logs POS tagging accuracy after the output tagged datasets have been written.

Dependency Parsing: Transition-Based Parsing

Now that we have a prediction for the grammatical role of the words, we want to understand how the words in the sentence relate to each other. This parser is built around the head-modifier construction: for each word, we choose a syntactic head that it modifies according to some grammatical role.

An example for the above sentence is as follows:


Below each word in the sentence we see both a fine-grained part-of-speech (PRP, VBD, DT, NN etc.), and a coarse-grained part-of-speech (PRON, VERB, DET, NOUN, etc.). Coarse-grained POS tags encode basic grammatical categories, while the fine-grained POS tags make further distinctions: for example NN is a singular noun (as opposed, for example, to NNS, which is a plural noun), and VBD is a past-tense verb. For more discussion see Petrov et al. (2012).

Crucially, we also see directed arcs signifying grammatical relationships between different words in the sentence. For example I is the subject of saw, as signified by the directed arc labeled nsubj between these words; man is the direct object (dobj) of saw; the preposition with modifies man with a prep relation, signifiying modification by a prepositional phrase; and so on. In addition the verb saw is identified as the root of the entire sentence.

Whenever we have a directed arc between two words, we refer to the word at the start of the arc as the head, and the word at the end of the arc as the modifier. For example we have one arc where the head is saw and the modifier is I, another where the head is saw and the modifier is man, and so on.

The grammatical relationships encoded in dependency structures are directly related to the underlying meaning of the sentence in question. They allow us to easily recover the answers to various questions, for example whom did I see?, who saw the man with glasses?, and so on.

SyntaxNet is a transition-based dependency parser Nivre (2007) that constructs a parse incrementally. Like the tagger, it processes words left-to-right. The words all start as unprocessed input, called the buffer. As words are encountered they are put onto a stack. At each step, the parser can do one of three things:

  1. SHIFT: Push another word onto the top of the stack, i.e. shifting one token from the buffer to the stack.
  2. LEFT_ARC: Pop the top two words from the stack. Attach the second to the first, creating an arc pointing to the left. Push the first word back on the stack.
  3. RIGHT_ARC: Pop the top two words from the stack. Attach the second to the first, creating an arc point to the right. Push the second word back on the stack.

At each step, we call the combination of the stack and the buffer the configuration of the parser. For the left and right actions, we also assign a dependency relation label to that arc. This process is visualized in the following animation for a short sentence:


Note that this parser is following a sequence of actions, called a derivation, to produce a "gold" tree labeled by a linguist. We can use this sequence of decisions to learn a classifier that takes a configuration and predicts the next action to take.

Training a Parser Step 1: Local Pretraining

As described in our paper, the first step in training the model is to pre-train using local decisions. In this phase, we use the gold dependency to guide the parser, and train a softmax layer to predict the correct action given these gold dependencies. This can be performed very efficiently, since the parser's decisions are all independent in this setting.

Once the tagged datasets are available, a locally normalized dependency parsing model can be trained with the following command:

bazel-bin/syntaxnet/parser_trainer \
  --arg_prefix=brain_parser \
  --batch_size=32 \
  --projectivize_training_set \
  --decay_steps=4400 \
  --graph_builder=greedy \
  --hidden_layer_sizes=200,200 \
  --learning_rate=0.08 \
  --momentum=0.85 \
  --output_path=models \
  --task_context=models/brain_pos/greedy/$PARAMS/context \
  --seed=4 \
  --training_corpus=tagged-training-corpus \
  --tuning_corpus=tagged-tuning-corpus \

Note that we point the trainer to the context corresponding to the POS tagger that we picked previously. This allows the parser to reuse the lexicons and the tagged datasets that were created in the previous steps. Processing data can be done similarly to how tagging was done above. For example if in this case we picked parameters 200x200-0.08-4400-0.85-4, the training, tuning and dev sets can be parsed with the following command:

for SET in training tuning dev; do
  bazel-bin/syntaxnet/parser_eval \
    --task_context=models/brain_parser/greedy/$PARAMS/context \
    --hidden_layer_sizes=200,200 \
    --input=tagged-$SET-corpus \
    --output=parsed-$SET-corpus \
    --arg_prefix=brain_parser \
    --graph_builder=greedy \

Training a Parser Step 2: Global Training

As we describe in the paper, there are several problems with the locally normalized models we just trained. The most important is the label-bias problem: the model doesn't learn what a good parse looks like, only what action to take given a history of gold decisions. This is because the scores are normalized locally using a softmax for each decision.

In the paper, we show how we can achieve much better results using a globally normalized model: in this model, the softmax scores are summed in log space, and the scores are not normalized until we reach a final decision. When the parser stops, the scores of each hypothesis are normalized against a small set of possible parses (in the case of this model, a beam size of 8). When training, we force the parser to stop during parsing when the gold derivation falls off the beam (a strategy known as early-updates).

We give a simplified view of how this training works for a garden path sentence, where it is important to maintain multiple hypotheses. A single mistake early on in parsing leads to a completely incorrect parse; after training, the model learns to prefer the second (correct) parse.

Beam search training

Parsey McParseface correctly parses this sentence. Even though the correct parse is initially ranked 4th out of multiple hypotheses, when the end of the garden path is reached, Parsey McParseface can recover due to the beam; using a larger beam will get a more accurate model, but it will be slower (we used beam 32 for the models in the paper).

Once you have the pre-trained locally normalized model, a globally normalized parsing model can now be trained with the following command:

bazel-bin/syntaxnet/parser_trainer \
  --arg_prefix=brain_parser \
  --batch_size=8 \
  --decay_steps=100 \
  --graph_builder=structured \
  --hidden_layer_sizes=200,200 \
  --learning_rate=0.02 \
  --momentum=0.9 \
  --output_path=models \
  --task_context=models/brain_parser/greedy/$PARAMS/context \
  --seed=0 \
  --training_corpus=projectivized-training-corpus \
  --tuning_corpus=tagged-tuning-corpus \
  --params=200x200-0.02-100-0.9-0 \
  --pretrained_params=models/brain_parser/greedy/$PARAMS/model \

Training a beam model with the structured builder will take a lot longer than the greedy training runs above, perhaps 3 or 4 times longer. Note once again that multiple restarts of training will yield the most reliable results. Evaluation can again be done with In this case we use parameters 200x200-0.02-100-0.9-0 to evaluate on the training, tuning and dev sets with the following command:

for SET in training tuning dev; do
  bazel-bin/syntaxnet/parser_eval \
    --task_context=models/brain_parser/structured/$PARAMS/context \
    --hidden_layer_sizes=200,200 \
    --input=tagged-$SET-corpus \
    --output=beam-parsed-$SET-corpus \
    --arg_prefix=brain_parser \
    --graph_builder=structured \

Hooray! You now have your very own cousin of Parsey McParseface, ready to go out and parse text in the wild.


To ask questions or report issues please post on Stack Overflow with the tag syntaxnet or open an issue on the tensorflow/models issues tracker. Please assign SyntaxNet issues to @calberti or @andorardo.


Original authors of the code in this package include (in alphabetical order):

  • Alessandro Presta
  • Aliaksei Severyn
  • Andy Golding
  • Bernd Bohnet
  • Chris Alberti
  • Daniel Andor
  • David Weiss
  • Emily Pitler
  • Greg Coppola
  • Ivan Bogatyy
  • Ji Ma
  • Keith Hall
  • Kuzman Ganchev
  • Livio Baldini Soares
  • Michael Collins
  • Michael Ringgaard
  • Ryan McDonald
  • Slav Petrov
  • Stefan Istrate
  • Terry Koo
  • Tim Credo