This repository releases the code for Generating Query Focused Summaries from Query-Free Resources.

Please cite the following paper [bib] if you use this code,

Xu, Yumo, and Mirella Lapata. "Generating Query Focused Summaries from Query-Free Resources." In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pp. 6096–6109. 2021.

The availability of large-scale datasets has driven the development of neural models that create generic summaries from single or multiple documents. In this work we consider query focused summarization (QFS), a task for which training data in the form of queries, documents, and summaries is not readily available. We propose to decompose QFS into (1) query modeling (i.e., finding supportive evidence within a set of documents for a query) and (2) conditional language modeling (i.e., summary generation). We introduce MaRGE, a Masked ROUGE Regression framework for evidence estimation and ranking which relies on a unified representation for summaries and queries, so that summaries in generic data can be converted into proxy queries for learning a query model. Experiments across QFS benchmarks and query types show that our model achieves state-of-the-art performance despite learning from weak supervision.

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

Preliminary setup

Project structure

└───log        # logging files
└───run        # scripts for MaRGE training
└───src        # source files
└───data       # generic data for training; qfs data for test/dev
└───graph      # graph components for query expansion
└───model      # MaRGE models for inference
└───rank       # ranking results
└───text       # summarization results
└───unilm_in   # input files to UniLM
└───unilm_out  # output files from UniLM

After cloning this project, use the following command to initialize the structure:

mkdir log data graph model rank text unilm_in unilm_out

Creating environment

cd ..
virtualenv -p python3.6 marge
cd marge
. bin/activate
pip install -r requirements.txt

You need to install apex:

cd ..
git clone
cd apex
python3 install

Also, you need to setup ROUGE evaluation if you have not yet done it. Please refer to this repository. After finishing the setup, specify the ROUGE path in frame/utils/ as an attribute of PathParser:

self.rouge_dir = '~/ROUGE-1.5.5/data'  # specify your ROUGE dir

Preparing benchmark data

Since we are not allowed to distribute DUC clusters and summaries, you can request DUC 2005-2007 from NIST. After acquiring the data, gather each year's clusters and summaries under data/duc_cluster and data/duc_summary, respectively. For instance, DUC 2006's clusters and summaries should be found under data/duc_cluster/2006/ and data/duc_summary/2006/, respectively. For DUC queries: you don't have to prepare queries by yourself; we have put 3 json files for DUC 2005-2007 under data/masked_query, which contain a raw query and a masked query for each cluster. Queries will be fetched from these files at test time.

TD-QFS data can be downloaded from here. You can also use the processed version here.

After data preparation, you should have the following directory structure with the right files under each folder:

│   └───duc_clusters   # DUC clusters 
│   └───duc_summaries  # DUC reference summaries 
│   └───masked_query   # DUC queries (raw and masked)
│   └───tdqfs          # TD-QFS clusters, queries and reference summaries

MaRGE: query modeling

Preparing training data

Source files for building training data are under src/sripts. For each dataset (Multi-News or CNN/DM), there are three steps create MaRGE training data.

A training sample for Marge can be represented as {sentence, masked summary}->ROUGE(sentence, summary). So we need to get the ROUGE scores for all sentences (step 1) and creating masked summaries (step 2). Then we put them together (step 3).

  1. Calculate ROUGE scores for all sentences:

    python src/sripts/

  2. Build masked summaries:

    python src/sripts/

  3. Build train/val/test datasets:

    python src/sripts/

In our experiments, Marge trained on data from Multi-News yielded the best performance in query modeling. If you want to build training data from CNN/DM:

  1. Use the function gathered_mp_dump_sentence_cnndm() in the first step (otherwise, use the function gathered_mp_dump_sentence_mn() )
  2. Set dataset='cnndm' in the second step (otherwise, dataset='mn')
  3. Use instead for the last step

Model training

Depending on which training data you have built, you can run either one of the following two scripts:

. ./run/   # train MaRGE with data from CNN/DM
. ./run/  # train MaRGE with data from Multi-News

Configs specified in these two files are used in our experiments, but feel free to change them for further experimentation.

Inference and evaluation

Use src/frame/rr/ for DUC evaluation and src/frame/rr/ for TD-QFS evalaution. We will take DUC evaluation for example.

In src/frame/rr/, run the following methods in order (or at once):

dump_rel_scores()  # inference with MaRGE
rel_scores2rank()  # turn sentence scores to sentence rank
rr_rank2records()  # take top sentences

To evaluate evidence rank, in src/frame/rr/, run:


MaRGESum: summary generation

Prepare training data from Multi-News

To train a controllable generator, we make the following three changes to the input from Multi-News (and CNN/DM):

  1. Re-order input sentences according to their ROUGE scores, so the top ones will be biased over:

    python scripts/

  2. Prepend a summary-length token

  3. Prepend a masked summary (UMR-S)

Prepare training data from CNN/DM

Our best generation result is obtained with CNN/DM data. To train MargeSum on CNN/DM data, apart from the above-mentioned three customizations, we need an extra step: build a multi-document version of CNN/DM.

This is mainly because the summaries in the original CNN/DM are fairly short, while testing on QFS requires 250 words as output. To fix this issue, we concatenate summaries from a couple of relevant samples to get a long enough summary. Therefore, the input is now a cluster of the documents from these relevant samples.

This involves in Dr.QA to index all summaries in CNN/DM. After indexing, you can use the following script to cluster samples via retrieving similar summaries:

python scripts/
  • upload the training data, so you can use this multi-document CNN/DM without making it from scratch.

Inference and evaluation

Setting up UniLM environment

To evaluate abstractive summarization, you need to setup an UniLM evironment following the instructions here.

After setting up UnILM, in src/frame/rr/, run:


This turns ranked evidence from Marge into MargeSum input files.

Now You can evaluate the trained UniLM model for developement and testing. Go to the UniLM project root, set the correct input directory, and deocode the summaries.

  • add detailed documentation for setting up UniLM.
  • add detailed documentation for decoding.

To evaluate the output, use the following function in src/frame/rr/


You can specifiy inference configs in src/frame/rr/