Question about acos task on the Chinese dataset

[twotwoiscute]

Hi,Thanks for the great work, I have a question on the task of acos
I've downloaded the dataset from here and there's a task called acos which I am interested in.

Observation

I try a quick train on all the dataset in acos_datasets and found out the fact that the model does not do well on dataset 505.Chinese_Zhang, the fine-tuned model always output the empty string even the value of both training ands val loss decrease during process.

So I check the paper and realize the dataset paper used is shown below:

Dataset: SemEval 2014,15 and 16 datasets are
used for our experimentation. The dataset is used
as a benchmark for ABSA tasks and has customer
reviews from three domains; laptops (Lapt14), hotels (Hotel15), and restaurants (Rest14, Rest15, and
Rest16)
which migh bring out the fact that the pretrained model it used in [training script] would not be so good at Chinese. (

PyABSA/pyabsa/tasks/ABSAInstruction/multitask_train.py

Line 21 in 192a528

model_checkpoint = "kevinscaria/ate_tk-instruct-base-def-pos-neg-neut-combined"

)

What I have done

I am new to NLP, I believe there's always the step that deal with tokenizer like the codeAutoTokenizer.from_pretrained before training and here's how I init the model, I downloaded it here directly.

self.tokenizer = AutoTokenizer.from_pretrained(checkpoint)
self.model = AutoModelForSeq2SeqLM.from_pretrained(checkpoint)

Question

Is it okay to apply the original pipeline that the model is trained on english I/O directly to the model that takes chinese input I/O? if not what should I do before get into training. Thanks!

[twotwoiscute]

I also try train dataset 505.Chinese_Zhang only, the training process looks to go me.
However when it comes to predict, the model still output empty string as described above.

[yangheng95]

Sorry for my late reply as I am on a vacation. Saddly, the ACOS task is only tested for English, while any other language is enabled, I will release a new version. BTW, Chinese is supported by ATEPC task.

[twotwoiscute]

Thanks for the reply , from your reply I can tell that you’ve known all it takes to get the things done. Can you tell me the pipeline of how to achieve it? I would like to try it on my own.Thanks!