Our method consists of two steps, where the first is to conduct monolingual summarization, and the second is to conduct crosslingual summarization.

For monolingual summarization, we use UniSumm, which is further fine-tuned on the few-shot monolingual in ConvSumX.

For crosslingual summarization, our method first involves a data-preprocessing step. The preprocessing is to replace original dialogue with one that is concated with monolingual summary, as en_summ [lsep] en_doc, where en_summ is the english (monolingual) summary, en_doc is the english dialogue, and [lsep] is a special token. Then, you can train the model using below command (we show a):

python mbart-ft.py \
        --model_path facebook/mbart-large-50-many-to-many-mmt \
        --tokenizer_path facebook/mbart-large-50-many-to-many-mmt \
        --data_root $data_path \
        --src_lang en_XX  --tgt_lang  fr_XX\
        --input_type "doc_en"  --output_type "sum_fr" \
        --save_prefix $prefix \
        --fp32 --device_id 1 \
        --batch_size 8 --epochs 20 \
        --lr $lr --grad_accum $acc --warmup $warm --dataset_size 157 \
        --val_batch_size 16 \
        --min_output_len 5 --max_output_len 50

The above is adapted from https://github.com/krystalan/ClidSum/, contributed by @Huajian Zhang.