Biased Random Walk based Social Regularization for Word Embeddings

Socialized word embeddings (SWE) has been proposed to deal with two phenomena of language use:

everyone has his/her own personal characteristics of language use,
and socially connected users are likely to use language in similar ways.

We observe that the spread of language use is transitive, namely, one user can affect his/her friends, and the friends can also affect their friends. However, SWE modeled transitivity implicitly. The social regularization in SWE only applies to one-hop friends, and thus users outside the one-hop social circle will not be affected directly.

In this work, we adopt random walk methods to generate paths on the social graph to model the transitivity explicitly. Each user on a path will be affected by his/her adjacent user(s) on the path. Moreover, according to the update mechanism of SWE, fewer friends a user has, fewer update opportunities he/she can get. Hence, we propose a biased random walk method to provide these users with more update opportunities.

Experiments show that our random walk based social regularizations perform better on sentiment classification task.

Preparation

You need to download the dataset:

Download Yelp dataset
Convert following datasets from json format to csv format by using json_to_csv_converter.py in preprocess and get two .json files:
- yelp_academic_dataset_review.json
- yelp_academic_dataset_user.json
And put them in the data directory.
Download LIBLINEAR or Multi-core LIBLINEAR and put the liblinear under the root directory of the repo. And you can install LIBLINEAR according to its Installation

Preprocessing

cd preprocess

Modify run.py by specifying --input (Path to yelp dataset).

python run.py

Training

cd train

You may modify the following arguments in run.py:

--yelp_round The round number of yelp data, e.g. {9, 10}
--para_lambda The trade off parameter between log-likelihood and regularization term
--para_r The constraint of the L2-norm
--para_path The number of random walk paths for every review
--para_path_length The length of random walk paths for every review
--path_p The return parameter for the second-order random walk
--path_q The in-out parameter for the second-order random walk
--para_alpha The restart parameter for the bias random walk
--para_bias_weight The bias parameter for the bias random walk

Because there are a number of ways to train models, you also need to specify model_types in training.py.

Then begin to train models:

python run.py

Sentiment Classification

cd sentiment

You may modify the following arguments in run.py, arguments are the same as above.

It will run two .py files, where sentiment.py is about sentiment classification on all users while head_tail.py is about sentiment classification on head users and tail users. You can choose one or both.

Because there are a number of ways to train models, you also need to specify model_types in sentiment.py and head_tail.py.

Then begin the classification:

python run.py

User vectors for attention

We thank Tao Lei as our code is developed based on his code.

cd attention

You can simply re-implement our results of different settings by modifying the run.sh:

add user and word embeddings by specifying --user_embs and --embedding.
add train/dev/test files by specifying --train, --dev, and --test respectively.
choose the type of layers by specifying --layer, cnn or lstm.
three settings for our experiments could be achieved by specifying --user_atten and --user_atten_base:
- setting --user_atten 0 for Without attention.
- setting --user_atten 1 --user_atten_base 1 for Trained attention.
- setting --user_atten 1 --user_atten_base 0 for Fixed user vector as attention.

Then begin to train attention models:

bash run.sh

Pretrained embeddings and models

You can download pretrained embeddings and models in release.

Put embs/* in the embs/ directory.
Put sentiment_models/* in the sentiment/models/ directory.
Put attention_models/* in the attention/models/ directory.

Directory Structure

.
├── attention
│   ├── models
│   ├── nn
│   │   ├── __init__.py
│   │   ├── advanced.py
│   │   ├── basic.py
│   │   ├── evaluation.py
│   │   ├── initialization.py
│   │   └── optimization.py
│   ├── utils
│   │   └── __init__.py
│   ├── dc.py
│   └── run.sh
├── data
├── embs
│   ├── users_sample.txt
│   └── words_sample.txt
├── liblinear
├── preprocess
│   ├── english.pickle
│   ├── english_stop.txt
│   ├── json_to_csv_converter.py
│   ├── preprocess.py
│   └── run.py
├── sentiment
│   ├── format_data
│   ├── models
│   ├── results
│   ├── get_SVM_format_swe.c
│   ├── get_SVM_format_w2v.c
│   ├── head_tail.py
│   ├── run.py
│   └── sentiment.py
├── train
│   ├── run.py
│   ├── swe.c
│   ├── swe_with_2nd_randomwalk.c
│   ├── swe_with_bias_randomwalk.c
│   ├── swe_with_deepwalk.c
│   ├── swe_with_node2vec.c
│   ├── swe_with_randomwalk.c
│   ├── training.py
│   └── w2v.c
├── LICENSE
└── README.md

OS and Dependencies

*nix operating systems or WSL
Python 2.7
gcc
Theano (>= 0.7 but < 1.0)
NumPy
gensim
PrettyTable
Pandas
ujson
NLTK

Citation

If you use this code, then please cite our IJCAI 2018 paper:

@inproceedings{ijcai2018-634,
  title     = {Biased Random Walk based Social Regularization for Word Embeddings},
  author    = {Ziqian Zeng and Xin Liu and Yangqiu Song},
  booktitle = {Proceedings of the Twenty-Seventh International Joint Conference on
               Artificial Intelligence, {IJCAI-18}},
  publisher = {International Joint Conferences on Artificial Intelligence Organization},             
  pages     = {4560--4566},
  year      = {2018},
  month     = {7},
  doi       = {10.24963/ijcai.2018/634},
  url       = {https://doi.org/10.24963/ijcai.2018/634},
}

License

Licensed under the MIT License.

haowenke/SRBRW