Code and dataset for 'Contrastive Aligned Joint Learning for Multilingual Summarization'
CALMS
├── fairseq # fairseq 0.9.0
├── myfairseq/examples/summarization # fairseq user-dir
│ ├── criterions # loss functions
│ ├── data # dataset
│ ├── models # nn model
│ └── tasks # task definition
├── preprocess # process from raw data
│ ├── checkOverlap.py # check overlap
│ ├── dataClear.py # clear raw data
│ └── dataSplit.py # split train/dev/test
├── RELEASE-1.5.5 # pyrouge dependencies
├── scripts # shell scripts for preprocess
│ ├── concat_lang5.sh # concat five languages
│ ├── filterdata.sh # filter raw data
│ └── predata.sh # tokenize dataset and create data-bin
├── utils # others
│ ├── trans # denoise preprocess
│ ├── calRouge.py # calculate multilingual rouge
│ ├── getbaseline.py # calculate lead and oracle
│ ├── logConfig.py # log class
│ ├── makeLabel.py # extract extractive label
│ └── tokenizer.py # tokenizer
├── finetune_joint.sh # jointly training multiple languages
├── finetune_lang5.sh # finetune a single language from the unified model
├── finetune.sh # finetune a single language from mBART
├── generate_zero.sh # generate from the unified model directly
├── rankSumm.sh # CSR: Contrastive Sentence Ranking
├── denoiseSumm.sh # SAS: Sentence Aligned Substitution
├── README.md
└── requirements.txt
We provide three different versions for various requirements:
- urls.tar.gz is the archive urls of MLGSum and the basic script to crawl. However, some urls are lost due to the revision.
- raw.tar.gz is the original content from the web pages. You can run scripts/filterdata.sh to process it and get MLGSum.
- clean.tar.gz is the final version of our MLGSum, which is reported in our paper. If you do not want to build the dataset by yourself, you can directly use this link.
Our code is mainly based on Python3 and Fairseq 0.9. You should first run
$ pip install fairseq==0.9
$ pip install -r requirements.txtto install the requirements. Besides, if you want to use pyrouge for evaluation, you need to check some other dependencies:
$ sudo apt-get install libxml-perl libxml-dom-perl
$ pip install git+git://github.com/bheinzerling/pyrougeThe detailed information for ROUGE can be found in MLROUGE.
Note that the above steps have been included in training shell scripts (eg. finetune.sh). You can run them directly.
You should tokenize and make bin files from MLGSum by yourself. We have provide the basic scripts (scripts/predata.sh) for you.
Pay attention, by default the mBART does not insert the language tags until the training phase. However, this makes it difficult for train multiple languages together. Thus, for the single language training, we follow the practice of mBART (summarization_from_pretrained_mbart), which does not insert the tags. And for the joint training, we insert the language tags during the data preprocess (using the extended dictionary file) and remove the insertion in training phrase (summarization_from_pretrained_wo_langtag).
Besides, for jointly training, the multiple languages should be mixed together. You can use the script (scripts/concat_lang5.sh) to concat and shuffle them. To conclude, your data-bin directory may look like this:
data-bin
├── wotag # for the single language baseline
│ ├── de
│ ├── en
│ ├── ru
│ ├── fr
│ ├── ...
│ └── zh
├── wtag # for jointly training with multiple languages
│ ├── lang5 # concat de+en+ru+fr+zh
│ ├── de
│ ├── en
│ ├── ru
│ ├── fr
│ ├── ...
└── └── zh
We build our model based on Fairseq. Thus, to run the code, you should use fairseq repository and put the myfairseq/examples/summarization under the fairseq/example directory. You can also change the '--user-dir' path in 'finetine*.sh' to avoid this operation.
The files in myfairseq/examples/summarization/data should be put under the fairseq/fairseq/data/ and the fairseq/fairseq/data/__init__.py must be modified to include them. (We have done this for you.)
We define several new tasks for our training strategies:
- summarization_from_pretrained_mbart: finetune the monolingual summarization directly from mBART model.
- summarization_from_pretrained_mbart_joint jointly finetune multilingual summarization from mBART model.
- summarization_from_pretrained_mbart_wo_tag add the language tags during the data proprecess and ignore them during the training.
- ranking_summaization [Contrastive Sentence Ranking] create the (positive, negative) pairs and rank them.
- denoise_summaization [Sentence Aligned Substitution] use the translated lead sentences and insert them into the original document.
- denoise_ranking_summaization first replace translated sentences and then create ranking pairs.
- ranking_denoise_summaization first create ranking pairs and then replace translated sentences.
You can use different training strategies by changing the --task settings in the command line of Fairseq. We also provide several training scripts to use (**.sh*). The usage of each script can be found in each file.
To generate summaries, you can use the summarization_from_pretrained_mbart for data without language tags and summarization_from_pretrained_mbart_wo_tag for data with language tags. You can also use generate mode in finetune.sh or generate_zero.sh for generation under different settings.
We modify the original ROUGE for multilingual summarization. It is the same as MLROUGE. We put a simple version under utils/calRouge.py .