Project Description
RNNSharp is a toolkit of recurrent neural network which is widely used for many different kinds of tasks, such as sequence labeling. It's written by C# language and based on .NET framework 4.0 or above version.

This page will introduces you about what is RNNSharp, how it works and how to use it. To get the latest source code, please visit [SOURCE CODE] page and download it by clicking "download" link. To get the demo package, please access [DOWNLOADS] page and download the package.


RNNSharp supports many different types of recurrent neural network (aka RNN) structures.In the aspect of historical memory, it supports BPTT and LSTM structures. And in respect of output layer structure, RNNSharp supports naive output layer and recurrent CRFs[1].

For BPTT and LSTM, BPTT-RNN is usually called as "simple RNN", since the structure of its hidden layer node is very simple. This type RNN is not good at preserving long time historical memory, but since the cost of its decoding is lower than LSTM, it can be easily applied on real product environment which needs high performance.

LSTM-RNN is more complex than BPTT-RNN, since its hidden layer node has inner-structure which helps it to save very long time historical memory. In general, LSTM has better performance than BPTT on longer sequences and its encoding process is easier to be paralleled than BPTT-RNN. However, since its decoding cost is much higher than BPTT, to apply it on real product environment directly, its decoding latency will be high than BPTT.

For naive output layer and recurrent CRFs. Actually, naive output layer means the standard RNN output layer. Many widely experiments and applications have proved that it's an excellent algorithm for online sequence labeling task, such as speech recognition, auto suggestion and so on. It has better performance than MMEM and the algorithms which use traditional N-gram features.

For recurrent CRFs (recurrent conditional random fields), it's a new type of CRF based on RNN. Compared with the above one, Recurrent-CRF can be used for many different types of sequence labeling tasks in offline, such as word segmentation, named entity recognition and so on. With the similar feature set, it has better performance than linear CRF, since the representation of its feature is richer than before.

Supported Feature Types

RNNSharp supports four types of feature set. They are template features, context template features, run time feature and word embedding features. These features are controlled by configuration file, the following paragraph will introduce what these features are and how to use them in configuration file.

Template Features

This type of features are generated by templates. By given templates, according corpus, the features will be generated. The template feature has binary-value. If the feature is existed in current token, its value will be as 1, otherwise, the value will be as 0. It's similar as CRFSharp features. In RNNSharp, TFeatureBin.exe is a console tool to generate this type of features.

In template file, each line describes one template which consists of prefix, id and rule-string. The prefix is used to indicate template type. So far, RNNSharp also supports U-type feature, so the prefix is always as "U". Id is used to distinguish different templates. And rule-string is used to guide CRFSharp to generate features. 

# Unigram

The rule-string has two types of form, one is constant string, and the other is macro. The simplest macro form is {“%x[row,col]”}. Row specifies the offset between current focusing token and generating feature token in row. Col specifies the absolute column position in corpus. Moreover, combined macro is also supported, for example: {“%x[row1, col1]/%x[row2, col2]”}. When generating feature set, macro will be replaced as specific string. A template file example as follows:


In this template file, assuming current focusing token is “York NNP E_LOCATION” in the first record in training corpus above, the generated unigram feature set as follows:

Although U07 and U08, U11 and U12’s rule-string are the same, we can still distinguish them by id string.

When building features, according templates, builder will generate feature set (like the example in above) from specific corpus and save them into feature set file.

In feature configuration file, keyword TFEATURE_FILENAME is used to specify the file name for template feature in binary format

Context Template Features

This type of features are based on template features and combined with context. Here is an example, if the configuration of this feature is "-1,0,1", then the generated feature will combine the features of current token with its previous token and next token. For instance, if the sentence is "how are you". the generated feature set will be {Feature("how"), Feature("are"), Feature("you")}.

In feature configuration file, keyword TFEATURE_CONTEXT is used to specify the tokens' context range for the feature.

Word Embedding Features

This type of features are used to describe the features of given token. It's very useful when we only have small labeled corpus, but have lots of unlabeled corpus. This type of feature is generated by WSDSharp. With lots of unlabeled corpus, in train mode, WSDSharp is able to generate vectors for each token. Note that, the token's granularity in both word embedding feature and training corpus for RNN model encoding should be aligned, otherwise, lots of tokens in training corpus will not be matched with the feature. For more detailed information about how to generate word embedding features, please visit WSDSharp homepage.

In RNNSharp, this feature also supports context feature. It will combine all features of given contexts into a single word embedding feature.

In feature configuration, it has three keywords: WORDEMBEDDING_FILENAME is used to specify the encoded word embedding data file name generated by WSDSharpWORDEMBEDDING_CONTEXT is used to specify the token's context range for this feature. And WORDEMBEDDING_COLUMN is used to specify the column index applied the feature in corpus

Run Time Features

Compared with other features generated offline, this feature is generated in run time. So far, this feature uses the result of previous tokens as run time feature for current token. 

In feature configuration, keyword RTFEATURE_CONTEXT is used to specify the context range of this feature.

Feature Configuration File

The file contain the configuration items for features. All of them have been introduced in above sections, and the following is an example. In console tool, use -ftrfile as parameter to specify feature configuration file

#The file name for template feature set

#The context range for template feature set. In below, the context is current token, next token and next after next token

#The word embedding data file name generated by WSDSharp

#The context range for word embedding. In below, the context is current token, previous token and next token

#The column index applied word embedding feature

#The run time feature

Training file format

Training corpus contains many records to describe what the model should be. For each record, it is split into one or many tokens and each token has one or many dimension features to describe itself. 

In training file, each record can be represented as a matrix and ends with an empty line. In the matrix, each row describes one token and its features, and each column represents a feature in one dimension. In entire training corpus, the number of column must be fixed.

When RNNSharp encodes, if the column size is N, according template file describes, the first N-1 columns will usually be used as input data to generate binary feature set and train model. The Nth column (aka last column) is the answer that the model should output. The means, for one record, if we have an ideal encoded model, given all tokens’ the first N-1 columns, the model should output each token’s Nth column data as answer.

There is an example (a bigger training example file is at download section, you can see and download it there): 

and CC S
are VBP S
major JJ S
financial JJ S
centers NNS S
p FW S
y NN S
h FW S
44 CD S

In above example, we designed output answer as "POS_TYPE". POS means the position of the term in the chunk or named entity, TYPE means the output type of the term. 

The example is for labeling named entities in records. It has two records and each token has three columns. The first column is the term of a token, the second column is the token’s pos-tag result and the third column is to describe whether the token is a named entity or a part of named entity and its type. The first and the second columns are input data for encoding model, and the third column is the model ideal output as answer.

For POS, it supports four types as follows: 
S : the chunk has only one term 
B: the begin term of the chunk 
M: one of the middle term in the chunk 
E: the end term of the chunk 

For TYPE, the example contains many types as follows: 
ORGANIZATION : the name of one organization 
LOCATION : the name of one location 
For output answer without TYPE, it's just a normal term, not a named entity. 

Test file format

Test file has the similar format as training file. The only different between training and test file is the last column. In test file, all columns are features for CRF model.

Tag Mapping File

This file is used to map tag name and its id. For readable, RNNSharp uses tag name in corpus, however, for high efficiency in encoding and decoding, RNNSharp maps tag name into a integer value. The mapping is defined in a file (-tagfile as parameter in console tool) and its format as follows: tag id \t tag name


Console Tool


RNNSharpConsole.exe is a console tool for recurrent neural network encoding and decoding. The tool supports two types of running mode. "train" mode is used for training model and "test" mode is used for predicting output tag from test corpus by given encoded model.

Encode Model

In this mode, the console tool can encode a RNN model by given feature set and training/validated corpus. The usage as follows:

RNNSharpConsole.exe -mode train <parameters>
Parameters for training RNN based model
-trainfile <string>: training corpus file
-validfile <string>: validated corpus for training
-modelfile <string>: encoded model file
-modeltype <int>: model structure: simple RNN is 0, LSTM-RNN is 1, default is 0
-ftrfile <string>: feature configuration file
-tagfile <string>: supported output tagid-name list file
-alpha <float>: learning rate, default is 0.1
-beta <float>: Regularization parameter, default is 1e-7
-bptt <int>: the step for back-propagation through time, default is 4
-layersize <int>: hidden layer size for training, default is 200
-crf <0/1>: training model by standard RNN(0) or RNN-CRF(1), default is 0
-maxiter <int>: maximum iteration for training, default is 20
-savestep <int>: save temporary model after every <int> sentence, default is 0

Example: RNNSharpConsole.exe -mode train -trainfile train.txt -validfile valid.txt -modelfile model.bin -tagfile tags.txt -layersize 200 -modeltype 0 -alpha 0.1 -bptt 4 -crf 1 -maxiter 20 -savestep 200K

Decode Model

In this mode, the console tool is used to predict output tags of given corpus. The usage as follows:

RNNSharpConsole.exe -mode test <parameters>
Parameters for predicting iTagId tag from given corpus
-testfile <string>: training corpus file
-modelfile <string>: encoded model file
-tagfile <string>: supported output tagid-name list file
-ftrfile <string>: feature configuration file
-outfile <string>: result output file

Example: RNNSharpConsole.exe -mode test -testfile test.txt -modelfile model.bin
-tagfile tags.txt -ftrfile features.txt -outfile result.txt


It's used to generate template feature set by given template and corpus files. For high performance accessing and save memory cost, the indexed feature set is built as double array in trie-tree by AdvUtils. The tool supports three modes as follows:

TFeatureBin.exe <parameters>
The tool is to generate template feature from corpus and index them into file
-mode <string> : support extract,index and build modes
   extract : extract features from corpus and save them as raw text feature list
   index : build indexed feature set from raw text feature list
   build : extract features from corpus and generate indexed feature set

Build mode

This mode is to extract features from given corpus according templates, and then build indexed feature set. The usage of this mode as follows:

TFeatureBin.exe -mode build <parameters>
This mode is to extract feature from corpus and generate indexed feature set
-template <string> : feature template file
-inputfile <string> : file used to generate features
-ftrfile <string> : generated indexed feature file
-minfreq <int> : min-frequency of feature

Example: TFeatureBin.exe -mode build -template template.txt -inputfile train.txt -ftrfile tfeature -minfreq 3

In above example, feature set is extracted from train.txt and build them into tfeature file as indexed feature set.

Extract mode

This mode is only to extract features from given corpus and save them into a raw text file. The different between build mode and extract mode is that extract mode builds feature set as raw text format, not indexed binary format. The usage of extract mode as follows:

TFeatureBin.exe -mode extract <parameters>
This mode is to extract features from corpus and save them as text feature list
-template <string> : feature template file
-inputfile <string> : file used to generate features
-ftrfile <string> : generated feature list file in raw text format
-minfreq <int> : min-frequency of feature

Example: TFeatureBin.exe -mode extract -template template.txt -inputfile train.txt -ftrfile features.txt -minfreq 3

In above example, according templates, feature set is extracted from train.txt and save them into features.txt as raw text format. The format of output raw text file is "feature string \t frequency in corpus". Here is a few examples:

U01:仲恺 \t 123
U01:仲文 \t 10
U01:仲秋 \t 12

U01:仲恺 is feature string and 123 is the frequency that this feature in corpus.

Index mode

This mode is only to build indexed feature set by given templates and feature set in raw text format. The usage of this mode as follows:

TFeatureBin.exe -mode index <parameters>
This mode is to build indexed feature set from raw text feature list
-template <string> : feature template file
-inputfile <string> : feature list in raw text format
-ftrfile <string> : indexed feature set

Example: TFeatureBin.exe -mode index -template template.txt -inputfile features.txt -ftrfile features.bin

In above example, according templates, the raw text feature set, features.txt, will be indexed as features.bin file in binary format.

Run on Linux/Mac

With Mono-project which is the third party .NET framework on Linux/Mac, RNNSharp is able to run on some non-Windows platforms without re-compile or modify, such as Linux, Mac and others.


The RNNSharp also provides some APIs for developers to leverage it into their projects. By download source code package and open RNNSharpConsole project, you will see how to use APIs in your project to encode and decode RNN models. Note that, before use RNNSharp APIs, you should add RNNSharp.dll as reference into your project.


  1.     Recurrent Conditional Random Field For Language Understanding
  2.     Recurrent Neural Networks for Language Understanding
  3.     RNNLM - Recurrent Neural Network Language Modeling Toolkit


Last edited Jul 23 at 1:47 PM by monkeyfu, version 18