What it does This repo bundles five implementations of abstractive text summarization—seq2seq with attention, pointer-generator networks, and reinforcement learning—into Jupyter notebooks designed for Google Colab. The author wires everything to Google Drive, so you don’t download datasets or rent hardware. It covers English, Hindi, Amharic, and Arabic, with evaluation metrics (BLEU, ROUGE variants) packaged in a custom “zaksum” format.

The interesting bit Most repos make you reconstruct environments and hunt for compatible TensorFlow versions. Here, each model is a single notebook with a working Colab link. The author also surfaces a real tension in the field: pointer-generators hybridize extractive and abstractive approaches, while RL tries to fix seq2seq’s habit of producing vague, factually loose summaries. The coverage mechanism is noted as still unfinished.

Key highlights

• Five models across three implementation tracks (attention, pointer-generator, RL)
• Pre-connected Google Drive integration for data and model storage
• Evaluation wrapper (zaksum_eval.ipynb) handling BLEU + four ROUGE variants
• Published academic work behind the collection (ICCES 2019, arXiv 2020)
• Companion blog series on Medium/HackerNoon walking through each architecture

Caveats

• Mixed Python 2.7 and 3.x across notebooks; pointer-generator and RL models specifically require 2.7
• Several models are direct modifications of existing repos (Currie, dongjun-Lee, abisee, yaserkl/RLSeq2Seq) with attribution but unclear delta
• “Arabic Summarization” and “soon isA Arabic” suggest some tracks are more aspirational than complete

Verdict Worth bookmarking if you’re teaching or learning seq2seq mechanics and want runnable baselines without setup friction. Skip if you need production-ready, maintained code—this is a tutorial collection, not a framework.