Weakly Supervised Aspect Extraction Using a Student-Teacher Co-Training Approach

---

Overview tasks:

• derive seed words using word vectors, NMF and clarity scoring function
• text pre-processing and word embedding fine-tunning on OPOSUM and organic dataset
• develop student-teacher co-training approach and verify correctness with OPOSUM
• apply the approach to organic dataset with three different word embedding models, i.e. word2vec, glove and BERT

Instructions

Running Experiments

conda create -n nlp python=3.7
conda activate nlp
pip install -r requirements.txt

Run parser.py with appropriate arguments.

Ex:

python parser.py --domain organic --experiment_mode once --freeze_emb 1 --lr 5e-4 --weight_decay 0.01 --dropout 0.4 \
    --batch_size 8 --inner_iter 1 --epochs 4 --pretrained glove --wv_path ../wv/glove_corpus_wotf1_wostw_pretrained.bin \
    --data_dir ../processed --general_asp 0 --num_aspect 6

• --domain -> organic, oposum, bags_and_cases etc
• --experiment_mode -> multi-times, once -> run the set of experiments once or multiple independent times
• --freeze_emb -> 1 to freeze embeddings. 0 otherwise.
• --pretrained -> glove, word2vec, bert-base-uncased
• --wv_path -> #path/file for word embeddings e.g. '../wv/w2v_corpus_wotf1_tuned.bin', or '../wv/glove_corpus_wotf1_wostw_pretrained.bin'
• --epochs -> number of epochs
• --inner_iter -> number of inner iterations
• --data_dir -> path for pre-processed data
• --general_asp -> index of general aspect of dataset -> 0 for organic
• --num_aspect -> number of aspects of interest -> 6 for organic

Sample runs with setup environments are available in scratch_experiments.ipynb Jupyter Notebook.

Jupyter Notebooks

Below is the description of available jupyter notebooks

• Clarity Scoring.ipynb: derive seed words with clarity scoring function

• Clarity_Scoring_coarse.ipynb: derive seed words with clarity scoring function using coarse attributes from annotated dataset

• draft_plot.ipynb: some functions for plotting errorbar

• extract_from_json.ipynb: extract useful content from organic dataset, clean up articles, output json files for follow tasks

• extractAnnotatedData.ipynb: create organic test set

• learn_transformers.ipynb: experience following the official tutorial

• NMF.ipynb: derive seed words using NMF

• oposum_build_corpus.ipynb: prepare data for fine-tuning word embeddings

• oposum_glove_finetune.ipynb: fine-tune glove in oposum dataset

• oposum_w2v_fine-tune.ipynb: fine-tune word2vec in oposum datset

• organic_create_json.ipynb: create organic training set, sentencize and remove sentences in test set

• organic_experiments.ipynb: run experiments on the whole approach and teacher evaluation in organic dataset

• process_oposum_dataset.ipynb: transform oposum dataset into integer index, save useful files for fast loading

• process_organic_dataset.ipynb: the same as above but apply to organic dataset

• WordEmbeddings.ipynb: text pre-processing and fine-tune word2vec on organic dataset

• WV_cluster.ipynb: apply k-means to fine-tuned word embeddings

• draft_UnofficialLeverage.ipynb: code debugging, data inspection, etc.

• change_trials: experiments on different ideas/variants to change the unofficial code

Folder Structure

.
├─annotated-dataset
│  ├─annotated_3rd_round
│  ├─...
│  └─...
├─curated-source-dataset
│  ├─english
│  └─german
├─glove.6B                      # glove pre-trained model
├─ISWD	                        # code from official repo, for faster accessibility
├─JupyterNotebook               # jupyter notebook scripts for various tasks, each with description inside
├─LeverageJustAFewKeywords      # code from unofficial repo, changed for our usage
├─oposum_scripts                # pre-processing code from summarizing opinions paper
├─output                        # intermediate results and other output files
├─presentation                  # presentation
├─processed                     # processed data and other useful files
├─scripts                       # scripts intended 
└─wv                            # word vectors for oposum and organic dataset

Note:

• code in ./LeverageJustAFewKeywords is originated from https://github.com/aqweteddy/LeverageJustAFewKeywords. Further information can be checked in README.md in the subfolder.
• folder ./ISWD is from official code https://github.com/gkaramanolakis/ISWD, added with comments