Code for Exploring Segment Representations for Neural Segmentation Models

This is the code base for our IJCAI 2016 paper.

Prerequisite

1. cmake (~2.8)
2. git (~1.8)
3. g++ (~4.6 for c++11 features, 4.8 is used in this paper)
4. boost (~1.57)

Compile

Execute the following command to compile.

git submodule init
git submodule update
./configure
make

You should find the following executable files:

• ./bin/crf: the neural CRF baseline
• ./bin/labeler: the neural classifier baselin
• ./bin/semi_crf: the neural semi-CRF with prediction on segment label (for Named entity recognition)
• ./bin/semi_crf2: the neural semi-CRF without prediction on segment label (for Chinese word segmentation)

Data format

Input format

.tag file

Used in ./bin/labeler and ./bin/crf. Same with CoNLL03 format. Instances are separated by empty line. Each word in one instance occupy one line. See ./data_sample/ner/ner.train.tag for NER example and ./data_sample/cws/cws.train.tag for CWS example.

.seg file

Used in ./bin/semi_crf and ./bin/semi_crf2. Each instance in one line with ||| separating words and segmentation. See ./data_sample/ner/ner.train.seg for NER example and ./data_sample/cws/cws.train.seg for CWS example.

input unit embedding file

In the same format with word2vec.

segment embedding file

Similar to the word2vec embedding format, but entry and its vector are separated by tab. Since each entry (segment) consists one or more input units. Surface strings of its units are separated by space. See ./data_sample/ner/ner.segemb.sample for named entity embedding example and `./data_sample/cws/cws.segemb.sample for Chinese word example.

Running

Replace the ./data_sample/ner/ner.{train|devel|test}.tag with CoNLL03 data to reproduce the NER result in the paper.

Baseline

NN-Labeler

Taking ner for example, execute to train a model on sample data.

./run/ner_nlabeler.train.sh

look for the model under root dir with name of ner_bilstm_${args}.${pid}.params and execute

./run/ner_nlabeler.test.sh ner_bilstm_${args}.${pid}.params

to perform test process.

NN-CRF

• ./run/ner_ncrf.train.sh
• ./run/ner_ncrf.test.sh crf_${args}.${pid}.params

Neural Semi-CRF

SRNN

• ./run/ner_nsemicrf_srnn.train.sh
• ./run/ner_nsemicrf_srnn.test.sh semi_crf_${args}.${pid}.params

SCONCATE

• ./run/ner_nsemicrf_sconcate.train.sh
• ./run/ner_nsemicrf_sconcate.test.sh semi_crf_${args}.${pid}.params

SRNN+seg-embed

With Fine Tuning

• ./run/ner_nsemicrf_srnn_seg_wft.train.sh
• ./run/ner_nsemicrf_srnn_seg_wft.test.sh semi_crf_${args}.${pid}.params

Without Fine Tuning

• ./run/ner_nsemicrf_srnn_seg_woft.train.sh
• ./run/ner_nsemicrf_srnn_seg_woft.test.sh semi_crf_${args}.${pid}.params

SCONCATE+seg-embed

With Fine Tuning

• ./run/ner_nsemicrf_sconcate_seg_wft.train.sh
• ./run/ner_nsemicrf_sconcate_seg_wft.test.sh semi_crf_${args}.${pid}.params

Without Fine Tuning

• ./run/ner_nsemicrf_sconcate_seg_woft.train.sh
• ./run/ner_nsemicrf_sconcate_seg_woft.test.sh semi_crf_${args}.${pid}.params

Get help

Use --help option in the executable binaries to get more help. Or write to Yijia Liu oneplus.lau@gmail.com.