SeNsER

Code of paper:

SeNsER: Learning Cross-Building Sensor Metadata Tagger.

Yang Jiao*, Jiacheng Li*, Jiaman Wu, Dezhi Hong, Rajesh Gupta and Jingbo Shang. EMNLP (Findings) 2020.

Please cite above paper if you think the code or data is helpful.

Quick Links

• SeNsER
  • Quick Links
  • Model Notes
  • Installation
    • Dependencies
  • Data
    • Task
    • Format
  • Embedding
  • Training
  • Evaluation
  • Contact

Model Notes

SeNsER learns a sensor metadata tagger for a new building based on its raw metadata and some existing fully annotated building. It leverages the commonality between different buildings: At the character level, it employs bidirectional neural language models to capture the shared underlying patterns between two buildings and thus regularizes the feature learning process; At the word level, it leverages as features the k-mers existing in the fully annotated building. During inference, we further incorporate the information obtained from sources such as Wikipedia as prior knowledge.

Installation

Dependencies

The code is written in Python 3.6.

pip3 install -r requirements.txt

cd Model/PythonUtils/StringMatch & python setup.py install 

Data

We mainly focus on data from building apm and ebu3b.

Task

Chunking:

"Data/apm-ebu3b-chunk": train and dev on apm ,test on ebu3b

"Data/ebu3b-apm-chunk": train and dev on ebu3b, test on apm

Tagging:

"Data/apm-ebu3b-tag": train and dev on apm ,test on ebu3b

"Data/ebu3b-apm-tag": train and dev on ebu3b, test on apm

Format

Character Label

Empty lines are used as separators between sequences.

a B-
p I-
& I-
m I-
. O-
r B-
m I-
- I-
4 I-
7 I-
3 I-
1 I-
. O-
d B-
m I-
p I-
r I-
- O-
p B-
o I-
s I-

Embedding

Words in building domain are special. We must train the specific embedding with raw data before we do the following NER model training. Raw data of two buildings is in "Embedding/sentences". You can use "Embedding/word2vec.py" for embedding training.

Training

SeNsER (train on ebu3b, test on apm)

python train_wc.py

SeNsER-Kmer (train on ebu3b, test on apm)

python train_wc.py --no_dict

SeNsER-Dict (train on ebu3b, test on apm)

python train_wc.py --word_emb_weight 0

Evaluation

Evaluation will generate the output of model to 'Model/model_output.txt'.

python eval_wc.py

Contact

If you have any questions, please contact Jiacheng Li ([email protected])