Oreo: Text Summarization with Oracle Expectation

This repository releases the code for Text Summarization with Oracle Expectation (paper link):

Extractive summarization produces summaries by identifying and concatenating the most important sentences in a document. Since most summarization datasets do not come with gold labels indicating whether document sentences are summary-worthy, different labeling algorithms have been proposed to extrapolate oracle extracts for model training. In this work, we identify two flaws with the widely used greedy labeling approach: it delivers suboptimal and deterministic oracles. To alleviate both issues, we propose a simple yet effective labeling algorithm that creates soft, expectation-based sentence labels. We define a new learning objective for extractive summarization which incorporates learning signals from multiple oracle summaries and prove it is equivalent to estimating the oracle expectation for each document sentence. Without any architectural modifications, the proposed labeling scheme achieves superior performance on a variety of summarization benchmarks across domains and languages, in both supervised and zero-shot settings.

Should you have any query please contact me at [email protected].

Build Summarization Data with Oreo Labels

First download raw summarization data:

• CNN/DM: download link
• XSum, MultiNews, Reddit and WikiHow: download link

Put the downloaded datasets under raw_data.

You can then build your summarization data with Oreo labels for BertSum as follows:

Step 1: beam search

This step runs beam search over document sentences, and dump beam results into a json file. See json_data/cnndm_sample.valid.beam_json for output examples.

Run the following command for split={train, val, test}to build beams for all dataset splits:

split=train && dataset=CNNDM && beam=256 && summary_size=3 && src_json_fn=${split}_${dataset}_bert.jsonl && dump_json_fn=cnndm_bert-beams_${beam}-steps_${summary_size}.${split}.beam_json && py src/labels/build_beam_json.py --task build_beam_json_from_bert --src ~/oreo/raw_data/$src_json_fn --save ~/oreo/json_data/$dump_json_fn --beam $beam --summary_size $summary_size 

You should apply different --beam and --summary_size to different datasets.

Step 2: build sentence labels

The step derives Oreo based on constructed beams and dump results into a json file. See json_data/cnndm_sample.valid.bert_json for output examples.

Run the following command for split={train, val, test}to build labels for all dataset splits:

split=train && oracle_dist=uniform && beam=256 && summary_size=3 && beam_json_fn=cnndm_bert-beams_${beam}-steps_${summary_size}.${split}.beam_json && py src/labels/build_bert_json.py --task build_bert_json --src ~/oreo/json_data/$beam_json_fn --oracle_dist ${oracle_dist} --store_hard_labels --oracle_dist_topk 16

• You should apply different --oracle_dist_topk to different datasets.
• To get Ormax (a bound-prevserving variant introduced in our paper), further specify --hyp2sent_pool max.

Step 3: shard Bert files

oracle_dist=uniform_top_16_oracle_dist && beam=256 && summary_size=3 && save_dir=cnndm_bert-beams_${beam}-steps_${summary_size}-${oracle_dist}-hard_and_soft && py src/labels/build_bert_json.py --task shard_bert_json --save ~/oreo/json_data/${save_dir}

Step 4: format to PyTorch files

oracle_dist=uniform_top_16_oracle_dist && beam=256 && summary_size=3 && dir_name=cnndm_bert-beams_${beam}-steps_${summary_size}-${oracle_dist}-hard_and_soft && bert_json_dir=~/oreo/json_data/${dir_name} && bert_data_dir=~/oreo/bert_data/${dir_name} && py src/preprocess.py -mode format_to_bert_with_precal_labels -raw_path ${bert_json_dir} -save_path ${bert_data_dir} -lower -n_cpus 1 -log_file ./logs/preprocess.log 

Model training and evaluation

We are currently cleaning the code for model training and evaluation, and will soon release the code for both supervised and zero-shot summarization experiments. We recommend you to check out the official implementation of BertSum and GSum, which we used for extractive and abstractive experiments, respectively. Note that for a fair comparison between different labeling schemes, we follow their standard training configuration without any additional hyper-parameter optimization (e.g., for our specific labeling scheme).