anirudhagar13/dl_title_generation

Deep learning based title generation of scientific paper using abstracts as input, employing encoder-decoder based architecture.

title_generation

Uses attention networks to generate titles of technical papers using abstracts

Data Sources

• ARXIV data
• NIPS data

Structure

• backend: contains scripts for training and evaluation of models
• data: contains data files from training data, models to results

Setting up develope env and running server

• python3 -m virtualenv env
• source "virtual env path activate"
• pip install -r requirements.txt
• python manage.py runserver
• Can go to localhost and access web UI for evaluating latest saved model

How to train

• source "virtual env path activate"
• cd title_generation/deeptitles
• python -m main.backend.train > main/data/results/"dd_mm_HH.log"
• one can always go to the utility.py file to tune the hyperparams

How to train in MSI

• to do

---

Update with each commit

References

• Code: PyTorch Attention models

Current model details

• GRU based encoder and attention decoder
• Hidden size=512, dropout=0.1, learning_rate=0.01 (SGD)
• Uses teacher-forcing mechanism during training for faster convergence

Tasks to do

• setup an evaluation metric for prediction
• identify benchmark
• try other architecturs to make prediction better
• try hyper parameter tuning to make prediction better

Experimentation (for logging progress)

Commit	Description
811b1d8	NIPS Dataset (Simple Encoder + Attn_Decoder)