Can MLLMs Understand the Deep Implication Behind Chinese Images?

🔥News

• [Oct. 2024]: Our paper has released on arXiv, check it out!

Introduction

As the capabilities of Multimodal Large Language Models (MLLMs) continue to improve, the need for higher-order capability evaluation of MLLMs is increasing. However, there is a lack of work evaluating MLLM for higher-order perception and understanding of Chinese visual content. To fill the gap, we introduce the Chinese Image Implication understanding Benchmark, CII-Bench, which aims to assess the higher-order perception and understanding capabilities of MLLMs for Chinese images. CII-Bench stands out in several ways compared to existing benchmarks. Firstly, to ensure the authenticity of the Chinese context, images in CII-Bench are sourced from the Chinese Internet and manually reviewed, with corresponding answers also manually crafted. Additionally, CII-Bench incorporates images that represent Chinese traditional culture, such as famous Chinese traditional paintings, which can deeply reflect the model's understanding of Chinese traditional culture.
Through extensive experiments on CII-Bench across multiple MLLMs, we have made significant findings:
Initially, a substantial gap is observed between the performance of MLLMs and humans on CII-Bench. The highest accuracy of MLLMs attains 64.4%, where as human accuracy averages 78.2%, peaking at an impressive 81.0%. Subsequently, MLLMs perform worse on Chinese traditional culture images, suggesting limitations in their ability to understand high-level semantics and lack a deep knowledge base of Chinese traditional culture. Finally, it is observed that most models exhibit enhanced accuracy when image emotion hints are incorporated into the prompts.
We believe that CII-Bench will enable MLLMs to gain a better understanding of Chinese semantics and Chinese-specific images, advancing the journey towards expert artificial general intelligence (AGI).

CII-Bench

We introduce the Chinese Image Implication Understanding Benchmark CII-Bench, a new benchmark measuring the higher-order perceptual, reasoning and comprehension abilities of MLLMs when presented with complex Chinese implication images. These images, including abstract artworks, comics and posters, possess visual implications that require an understanding of visual details and reasoning ability. CII-Bench reveals whether current MLLMs, leveraging their inherent comprehension abilities, can accurately decode the metaphors embedded within the complex and abstract information presented in these images.

CII-Bench contains a total of 698 various Chinese images. These images are manually collected and annotated by 30 undergraduate students from various disciplines and institutions, with sources from multiple renowned Chinese illustration websites. Each image is manually designed with one to three multiple-choice questions, each with six options and only one correct answer. The questions cover the metaphors, symbolism, and detailed understanding of the images. The benchmark includes a total of 800 multiple-choice questions, with 765 questions used to construct the test set and 35 questions used to construct the development and validation set for few-shot tasks.

Evaluation

To obtain the model's score on CII-Bench, we need to first infer the model's predictions and then calculate the score.

1. Inference

Before inference, you need to configure the inference framework and code required for the model correctly. Below, we provide running examples for different inference frameworks.

Using lmdeploy for inference

First, you need to configure the environment required for running lmdeploy by following the steps in this link.

conda create -n lmdeploy python=3.8 -y
conda activate lmdeploy
pip install lmdeploy

Then, you need to configure the model configs in src/infer/models/__init__.py. Taking InternVL2-8B as an example, we pass ('.lmdeploy_chat', 'load_model') and ('.lmdeploy_chat', 'infer') to the load and infer fields respectively, indicating that the program will execute load and infer in src/infer/models/lmdeploy_chat to load the model and perform inference. If your model is stored locally, pass the model path to the model_path_or_name parameter and set call_type to local. Set tp according to your tensor parallelism requirements.

'InternVL2-8B': {
        'load': ('.lmdeploy_chat', 'load_model'),
        'infer': ('.lmdeploy_chat', 'infer'),
        'model_path_or_name': '/path/to/InternVL2-8B',
        'call_type': 'local',
        'tp': 1
    }

Run inference:

python infer/infer.py --config config/config_cii.yaml --split CII --mode none --model_name InternVL2-8B --output_dir results_cii --batch_size 4

--config: File field configuration
--split: Data split to run. When set to CII, it will infer both test and dev sets by default. You can modify the corresponding code in infer.py to run different splits.
--mode: Experiment setting. Default 'none' for zero-shot prediction. Options include none cot domain emotion rhetoric.
--model_name: Model name configured in src/infer/models/__init__.py.
--output_dir: Prediction output directory.
--batch_size: Inference batch size.

Similarly, you can use lmdeploy via the API server.

Launch the API service:

lmdeploy serve api_server OpenGVLab/InternVL2-8B --tp 4 --cache-max-entry-count 0.90 --backend turbomind --server-port 23333 --max-batch-size 1

Modify the model configuration to:

'InternVL2-8B': {
        'load': ('.api', 'load_model'),
        'infer': ('.api', 'infer'),
        'base_url': 'http://localhost:23333/v1',
        'api_key': 'YOUR_API_KEY',
        'model': 'OpenGVLab/InternVL2-8B',
        'call_type': 'api'
    }

And pass in the correct base_url, api_key, and model.

Now replace --batch_size with --num_workers, and assign the number of concurrent workers for inference.

Run inference:

python infer/infer.py --config config/config_cii.yaml --split CII --mode none --model_name InternVL2-8B --output_dir results_cii --num_workers 4

vLLM

First, configure the environment required for vLLM link

pip install vllm 

We recommend using the API service to call the model. Here we take Qwen2-VL-7B as an example.

Launch the API service:

vllm serve Qwen/Qwen2-VL-7B-Instruct --max-model-len 20000 --trust-remote-code --limit-mm-per-prompt image=4 --gpu-memory-utilization 0.9 --tensor-parallel-size 4 --port 8000

Modify the model configuration to:

'Qwen2-VL-7B': {
        'load': ('.api', 'load_model'),
        'infer': ('.api', 'infer'),
        'base_url': 'http://localhost:8000/v1',
        'api_key': 'YOUR_API_KEY',
        'model': 'Qwen/Qwen2-VL-7B-Instruct',
        'call_type': 'api'
    }

Run inference:

python infer/infer.py --config config/config_cii.yaml --split CII --mode none --model_name Qwen2-VL-7B --output_dir results_cii --num_workers 16

Custom Model

For models that are not supported by inference frameworks, you can manually add model inference files in the src/infer/models directory, such as idefics2.py. After adding the inference code and modifying the model configs, execute the following script to run inference.

Run inference:

python infer/infer.py --config config/config_cii.yaml --split CII --mode none --model_name idefics2-8b --output_dir results_cii --batch_size 4

2. Calculate Scores

Once the inference results are saved, directly execute eval_cii.py and eval_cii_sub.py in the src/ directory.

python eval_cii.py --evaluate_all --output_dir "results_cii" --save_dir "results_cii_with_status"
python eval_cii_sub.py --input_dir "results_cii_with_status"

Chinese Traditional Culture Evaluation

We choose to deeply analyze MLLM’s understanding of Chinese traditional culture by evaluating Chinese traditional paintings. For the detailed description and code, please refer to CTC_Evaluation.

Citation

If you find our work helpful in your research, please cite the following paper:

@misc{zhang2024mllmsunderstanddeepimplication,
      title={Can MLLMs Understand the Deep Implication Behind Chinese Images?}, 
      author={Chenhao Zhang and Xi Feng and Yuelin Bai and Xinrun Du and Jinchang Hou and Kaixin Deng and Guangzeng Han and Qinrui Li and Bingli Wang and Jiaheng Liu and Xingwei Qu and Yifei Zhang and Qixuan Zhao and Yiming Liang and Ziqiang Liu and Feiteng Fang and Min Yang and Wenhao Huang and Chenghua Lin and Ge Zhang and Shiwen Ni},
      year={2024},
      eprint={2410.13854},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2410.13854}, 
}