Automatic Generation of Topic labels

This repository contains the source code and data used for the paper:

Automatic Generation of Topic Labels (2020) Areej Alokaili, Nikolaos Aletras and Mark Stevenson in Proceedings of the 43rd International ACM SIGIR Conference on Research and Development in Information Retrieval(SIGIR ’20), July 25–30, 2020, Virtual Event, China. https://doi.org/10.1145/3397271.3401185 Pre-print

(A) Install required libraries

Python 3.6.9 is used.

1. TensorFlow V2
2. NumPy
3. scikit-learn
4. ipykernel

below libraries needed for evaluation only. You can skip if you want to do different evaluation metric other than BERTScore

1. sentencepiece
2. transformers
3. bert-score
4. matplotlib
5. pandas

use pip install -r requirements.txt to install all needed libraries

(B) Training

To run the model (data are processed and ready, only training is needed):

• Navigate to topic_labelling/

  1. To train the model with [inputs=top-30 terms from wikipedia article and outputs=wikipedia titles]

     python train_tf.py -m 'bigru_bahdanau_attention'  -d 'wiki_tfidf'

  2. To train the model with [inputs=first-30 words from wikipedia article and outputs=wikipedia titles] (refer to paper for details).

     python train_tf.py -m 'bigru_bahdanau_attention'  -d 'wiki_sent'

• Training will stop if no improvment is recorded and all checkpoints will be saved in training_checkpoint/data_name/ .

• Training options are detailed in the code or run

  python train_tf.py -h

(C) Inference (generate titles/labels)

1. Generate TITLES for a subset of wikipedia articles (1000 articles)

   python test_tf.py -m 'bigru_bahdanau_attention' -s 1000 -d 'wiki_tfidf' --load 'NAME_OF_CHECKPOINT'

   *replace NAME_OF_CHECKPOINT with the name of your checkpoint. For example, python test_tf.py -d 'wiki_tfidf' -m 'bigru_bahdanau_attention' --load bigru_bahdanau_attention_e_1_valloss2.19-2

2. Generate LABELS for bhatia_topics python test_tf.py -m 'bigru_bahdanau_attention' -s 1000 -d 'wiki_tfidf' --load 'NAME_OF_CHECKPOINT' -te 'bhatia_topics'

3. Generate LABELS for bhatia_topics_tfidf python test_tf.py -m 'bigru_bahdanau_attention' -s 1000 -d 'wiki_tfidf' --load 'NAME_OF_CHECKPOINT' -te 'bhatia_topics_tfidf'

4. Predictions, golds, and topics will be stored at results/data_name/ as

   • [model_name]_pred.out
   • [model_name]_gold.out
   • [model_name]_topics.out.

     (D) Evaluation

5. To measure the similarity between predicted and gold labels, python compute_bertscore.py -g results/path_to_gold_file.out -p results/path_to_predict_file.out

6. Output includes precision (P), recall (R) and f-score (F).

Repository hierarchy

• train_tf.py code to train the labelling network.

• test_tf.py code to generate new titles/labels.

• model_archi_tf.py neural network structure defind here.

• support_methods.py contain some method needed methods through out the system.

• extract_additional_terms_for_topics.ipynb notebook showing the steps taken to filter topic/labels pairs based on the overall human rating and matching them to similar documents to extract additional terms for bhatia_topics_tfidf.

• compute_bertscore.py: script to compute pairwise BERTScore between predicted titles/labels and gold titles/labels.

• data

  1. wiki_tfidf: contain files after preprocessing that are ready to be passed to the model.
  2. wiki_sent: contain files after preprocessing that are ready to be passed to the model.
  3. bhatia_topics: contains a csv file with two columns (column1: topic labels, columns2: topic's top 10 terms).
  4. bhatia_topics_tfidf: contains a csv file with two columns (column1: topic labels, columns2: topic's top 10 terms +20 terms from similar document (the 20 terms are extract using file extract_additional_topic_terms.ipynb).

• results: this is where the model's output are saved in text files.

• training_checkpoints: model checkpoints are saved here.