SENNA is a software distributed under a non-commercial license, which outputs a host of Natural Language Processing (NLP) predictions: part-of-speech (POS) tags, chunking (CHK), name entity recognition (NER), semantic role labeling (SRL) and syntactic parsing (PSG).
SENNA is fast because it uses a simple architecture, self-contained because it does not rely on the output of existing NLP system, and accurate because it offers state-of-the-art or near state-of-the-art performance.
SENNA is written in ANSI C, with about 3500 lines of code. It requires about 200MB of RAM and should run on any IEEE floating point computer.
Here are the main changes compared to SENNA v2.0:
-offsettags, which outputs start/end offsets (in the sentence) of each token.
DISCLAIMER: Our word embeddings differ from Joseph Turian's embeddings (even though it is unfortunate they have been called "Collobert & Weston embeddings" in several papers). Our embeddings have been trained for about 2 months, over Wikipedia.
SENNA's details concerning POS, CHK, NER and SRL tasks are included in a JMLR paper. Later, the techniques have been extended and applied to syntactic parsing (PSG), and published in a AISTATS paper. If you use SENNA, please consider citing these appropriate papers.
R. Collobert, J. Weston, L. Bottou, M. Karlen, K. Kavukcuoglu and P. Kuksa. Natural Language Processing (Almost) from Scratch, Journal of Machine Learning Research (JMLR), 2011.
R. Collobert. Deep Learning for Efficient Discriminative Parsing, in International Conference on Artificial Intelligence and Statistics (AISTATS), 2011.
We also provide binaries for a couple of platforms, in the same archive. The name of the executable for each platform is given as follow:
Everything is included in a single tar-gzipped file (185MB). Proceed to the download page.
gcc -o senna -O3 -ffast-math *.cYou might want to add additional suitable optimization flags for your platform. SENNA also compiles fine with the Intel compiler (icc).
If speed is critical, we recommend to compile SENNA with the Intel MKL
library, which provides a very efficient BLAS. Add the definition
as well as correct MKL libraries and include path.
gcc -o senna -O3 -ffast-math *.c -DUSE_MKL_BLAS [...]
SENNA also compiles with ATLAS BLAS. On our platform, the handcrafted code compiled with the gcc command line shown above was faster. However, if you want to use it, you can compile it with:
gcc -o senna -O3 -ffast-math *.c -DUSE_ATLAS_BLAS [...]
gcc -o senna -O3 -ffast-math *.c -DUSE_APPLE_BLAS -framework AccelerateThis will compile against Apple BLAS libraries included in your system. As for Linux, it is recommended to use Intel MKL library instead of Apple BLAS libraries. The following command line can be invoked (replacing the dots by the correct library and include paths):
gcc -o senna -O3 -ffast-math *.c -DUSE_MKL_BLAS [...]
Win32console project under Microsoft Visual Studio (you can download the Express Edition). Add all the includes and C file into the project, and build the solution. You can also use the command line (after opening a Visual Studio Command Prompt) in the following way:
cl /O2 /Fesenna.exe *.c
We recommend to use Intel MKL for speed. See your MKL manual for adding proper libraries and includes. Add also
the preprocessor definition
USE_MKL_BLAS in the project. Using the command line, it would be:
cl /O2 /Fesenna.exe *.c /DUSE_MKL_BLAS [...]
senna < sanity-test-input.txt > sanity-test-result.txtSENNA should create a file
sanity-test-result.txtwhich should be identical to the provided
sanity-test-input.txt comes from
the CoNLL 2000
chunking testing set. SENNA will output all tags for this file. It
should run in about 90 seconds on a decent computer (using MKL).
senna [options] < input.txt > output.txtOf course you can run SENNA in an interactive mode without the "pipes"
Each input line is considered as a sentence. SENNA has its own
tokenizer for separating words, which can be deactivated with
SENNA outputs one line per "token", with all the corresponding tags (in IOBES format) on the same line. An empty line is inserted between each output sentence. The first column is the token. Tags for all task then follow by default (POS, CHK, NER and SRL). Tags for SRL are preceded by a column which indicates if SENNA considered the token as a SRL verb or not ("-"). Then, there is one column per SRL verb.
SENNA supports the following options:
SENNA does not handle
... tokens. Please,
replace these tokens in your input text by the appropriate
(, ), .... Not replacing these tokens will have an impact on
performance (for e.g., POS accuracy goes down, from
We report here SENNA performance in per-word accuracy for POS, and F1 score
for all the other tasks. Timing corresponds to the time needed by SENNA to
pass over the given test data set (Macbook Pro i7, 2.8GHz, Intel MKL). For PSG, F1 score is the one over all
sentences (for sentences with less than 40 words, we get 88.5%).
|Part of Speech (POS)||(Toutanova et al, 2003)||(Accuracy)||97.29%||3|
|Chunking (CHK)||CoNLL 2000||(F1)||94.32%||2|
|Name Entity Recognition (NER)||CoNLL 2003||(F1)||89.59%||2|
|Semantic Role Labeling (SRL)||CoNLL 2005||(F1)||75.49%||36|
|Syntactic Parsing (PSG)||Penn Treebank||(F1)||87.92%||74|
Please email to Ronan Collobert for any problem report or positive feedback. We will be glad to hear from you.