Cross-lingual summarization (cross-sum) is a task to summarize a given document written in one language to another language short summary.

Traditional cross-sum approaches are based on two techniques namely early translation technique and late translation technique. Early translation can be explained easily as translate-then-summarize method. Late translation, in reverse, is summarize-then-translate method.

However, classical cross-sum methods tend to carry errors from monolingual summarization process or translation process to final cross-language output summary. Several end-to-end approaches have been proposed to tackle problems of traditional ones. Couple of end-to-end models are available to download as well.

💡Important Note In contrast to Zhu, et al, in our experiment, we found that filtering out articles using RTT technique worsen the overall performance of the end-to-end models significantly. Therefore, full datasets are highly recommended.

We used TR-TPBS as source documents for creating cross-lingual summarization dataset. In the same way as Zhu, et al., we constructed Th2En and Th2Zh by translating the summary references into target languages using translation service and filtered out those poorly-translated summaries using round-trip translation technique (RTT). The overview of cross-lingual summarization dataset construction is presented in belowe figure. Please refer to the corresponding paper for more details on RTT.

In our experiment, we set 𝑇1 and 𝑇2 equal to 0.45 and 0.2 respectively, backtranslation technique filtered out 27.98% from Th2En and 56.79% documents from Th2Zh.

python3 src/tools/cls_dataset_construction.py \
--dataset th2en \
--input_csv path/to/full_dataset.csv \
--output_csv path/to/save/filtered_csv \
--r1 0.45 \
--r2 0.2

• --dataset can be {th2en, th2zh}.
• --r1 and --r2 are where you can set ROUGE score thresholds (r1 and r2 represent ROUGE-1 and ROUGE-2 respectively) for filtering (assumingly) poor translated articles.

Click the file name to download.

Please refer to th2enzh_data_exploration.ipynb for more details.

• {th/en/zh}_title are only available in test set.
• {en/zh}_gg_sum are also only available in test set. We (at the time this experiment took place) assumed that Google translation was better than the main translation service we were using. We intended to use these Google translated summaries as some kind of alternative summary references, but in the end, they never been used. We decided to make them available in the test set anyway, just in case the others find them useful.
• {en/zh}_body were not presented during training end-to-end models. They were used only in early translation methods.

To evaluate these trained models, please refer to xls_model_evaluation.ipynb and ncls_model_evaluation.ipynb.

If you wish to evaluate the models with our test sets, you can use below script to create test files for XLS and NCLS models.

python3 src/tools/create_cls_test_manifest.py \
--test_csv_path path/to/testset.csv \
--output_dir path/to/save/testset_files \
--use_google_sum {true/false} \
--max_tokens 500 \
--create_ms_ref {true/false}

• output_dir is path to directory that you want to save test set files
• use_google_sum can be {true/false}. If true, it will select summary reference from columns {en/zh}_gg_sum. Default is false.
• max_tokens number of maximum words in input articles. Default is 500 words. Too short or too long articles can significantly worsen performance of the models.
• create_ms_ref whether to create Thai summary reference file to evaluate MS task in NCLS:CLS+MS model.

This script will produce three files namely test.CLS.source.thai.txt and test.CLS.target.{en/zh}.txt. test.CLS.source.thai.txt is used as a test file for cls task. test.CLS.target.{en/zh}.txt are the crosslingual summary reference for English and Chinese, they are used to evaluate ROUGE and BertScore. Each line is corresponding to the body articles in test.CLS.source.thai.txt.

🥳 We also evaluated MT tasks in XLS and NCLS:CLS+MT models. Please refers to xls_model_evaluation.ipynb and ncls_model_evaluation.ipynb for BLUE score results . For test sets that we used to evaluate MT task, please refer to data/README.md.

In this experiment, we used two automatic evaluation matrices namely ROUGE and BertScore to assess the performance of CLS models. We evaluated ROUGE on Chinese text at word-level, NOT character level.

We only reported BertScore on abstractive summarization models. To evaluate the results with BertScore we used weights from ‘roberta-large’ and ‘bert-base-chinese’ pretrained models for Th2En and Th2Zh respectively.

@mastersthesis{chumpolsathien_2020, 
    title={Using Knowledge Distillation from Keyword Extraction to Improve the Informativeness of Neural Cross-lingual Summarization},
    author={Chumpolsathien, Nakhun}, 
    year={2020}, 
    school={Beijing Institute of Technology}

• These cross-lingual datasets and the experiments are parts of Nakhun Chumpolsathien ’s master’s thesis at school of computer science, Beijing Institute of Technology. Therefore, as well, a great appreciation goes to his supervisor, Assoc. Prof. Gao Yang.
• Shout out to Tanachat Arayachutinan for the initial data processing and for introducing me to 麻辣烫, 黄焖鸡.
• We would like to thank Beijing Engineering Research Center of High Volume Language Information Processing and Cloud Computing Applications for providing computing resources to conduct the experiment.
• In this experiment, we used PyThaiNLP v. 2.2.4 to tokenize (on both word & sentence levels) Thai texts. For Chinese and English segmentation, we used Stanza.