The Framework PEARL

Environment

Computational platform: PyTorch 1.4.0, NVIDIA Geforce GTX 3090 (GPU), Inter i9-10900X (CPU), CUDA Toolkit 10.0

Development language: Python 3.6/C++

Liabraries are listed as follow, which can be installed via the command pip install -r requirements.txt.

numpy, scipy, tqdm, scikit-learn, sentencepiece=0.1.91, transformers, tensorboardX, nltk, os, sys, collections, itertools, argparse, subprocess, pickle, cudatoolkit=10.0, pytorch==1.4.0

Data sets

We provide all the data sets (profession data set, hobby data set, and 20News data set) in the folder data/datasets/.

Personal attribute prediction task

Profession data set(obtained from the authors of [2])
atribute values: 71; user utterances: 5747
used by the previous work: CHARM DSCGN

Hobby data set (obtained from the authors of [2])
atribute values: 149; user utterances: 5787
used by the previous work: CHARM DSCGN

Note that we follow the same task setting as previous personal attribute prediction papers[2-4], where attribute values are NOT explicitly mentioned in utterances and the given candidate attribute values are ranked based on the underlying semantics of utterances.

Weakly supervised text classification task

20News data set(obtained from [1])
classes: 5; documents: 17871
used by the previous work: X-Class

Note that PEARL is tested on the weakly supervised text classification task to verify its universality, flexibility and effectiveness.

Reproduce

Preprocess the profession data set:

CUDA_VISIBLE_DEVICES = [gpu_id] python static_representations.py --dataset_name profession
CUDA_VISIBLE_DEVICES = [gpu_id] python utterance_word_representations.py --dataset_name profesion

Similarly, the hobby (resp. 20News) data set can be preprocessed by replacing "profession" as "hobby" (resp. "20News").

Run PEARL on the profession data set:

python iterate_frame_profession.py

Similarly, PEARL can run on the hobby (resp. 20News) data set via the command "python iterate_frame_hobby.py" (resp. "python iterate_frame_20News.py").

[1] Lang K. Newsweeder. Learning to filter netnews. Machine Learning Proceedings 1995, 331-339.

[2] Tigunova A, Yates A, Mirza P, et al. CHARM: Inferring personal attributes from conversations. EMNLP'20, 5391-5404.

[3] Liu Y, Chen H, Shen W. Personal Attribute Prediction from Conversations. WWW'2022, 223-227.

[4] Tigunova A, Yates A, Mirza P, et al. Listening between the lines: Learning personal attributes from conversations. WWW'2019, 1818-1828.