LoTLIP: Improving Language-Image Pre-training for Long Text Understanding

Wei Wu1, Kecheng Zheng2,3, Shuailei Ma4, Fan Lu1, Yuxin Guo5, Yifei Zhang6, Wei Chen3, Qingpei Guo2, Yujun Shen2, Zheng-Jun Zha1

1University of Science and Technology of China, 2Ant Group, 3Zhejiang University, 4Northeastern University, China, 5Institute of Automation, Chinese Academy of Sciences, 6Shanghai Jiao Tong University

Understanding long text is of great demands in practice but beyond the reach of most language-image pre-training (LIP) models. In this work, we empirically confirm that the key reason causing such an issue is that the training images are usually paired with short captions, leaving certain tokens easily overshadowed by salient tokens. Towards this problem, our initial attempt is to relabel the data with long captions, however, directly learning with which may lead to performance degradation in understanding short text (e.g., in the image classification task). Then, after incorporating corner tokens to aggregate diverse textual information, we manage to help the model catch up to its original level of short text understanding yet greatly enhance its capability of long text understanding. We further look into whether the model can continuously benefit from longer captions and notice a clear trade-off between the performance and the efficiency. Finally, we validate the effectiveness of our approach using a self-constructed large-scale dataset, which consists of 100M long caption oriented text-image pairs. It is noteworthy that, on the task of long-text image retrieval, we beat the competitor using long captions with 11.1% improvement (i.e., from 72.62% to 83.72%). We will release the code, the model, and the new dataset to facilitate the reproducibility and further research.

Paper Code Dataset (coming soon)

Dataset

Our dataset contains long captions for 100M images, where each image has three paired long captions generated by InstructBLIP, LLAVA, and ShareGPT4V, respectively.

In the image, a person's finger is seen pointing towards a cat's nose. The cat, with its fur in shades of gray and white, is looking up at the finger with its mouth slightly open, as if curious or interested in the action. The background is a plain white wall, providing a stark contrast to the subjects in the foreground. The cat's position relative to the finger suggests it's being the center of attention, possibly for a playful interaction or a trick. The image captures a moment of connection between the person and the cat, highlighting the curiosity and engagement of the feline.

The image captures a lively scene at a beach. The beach is nestled against a rocky cliff, which is adorned with patches of green vegetation. The water surrounding the beach is a stunning turquoise color, reflecting the clear blue sky above. There are several people enjoying the beach, some are frolicking in the water while others are sunbathing on the sand. The beach is not just a place for relaxation but also a place for adventure, as evidenced by the presence of a few boats in the water.The image is taken from a high vantage point, providing a panoramic view of the beach and the surrounding landscape. This perspective allows us to appreciate the vastness of the beach and the expansive water body. The relative positions of the objects suggest a spacious and uncrowded beach, with ample room for everyone to enjoy their day.

The image features a white and pink double-decker bus traveling down a city street. The bus has two levels, making it a double-decker vehicle. It is parked at the side of the road, likely waiting for passengers to board or disembark. In the background, there are several other vehicles, including cars, trucks, and motorcycles, adding to the bustling atmosphere of the urban setting.'

Evaluation Dataset

Existing short-text-image retrieval tasks primarily rely on short textual input. We collected long text-image pairs from DCI, IIW, and ShareGPT4V datasets for constructing long-text-image retrieval evaluation task.

Benefits and Drawbacks from Long Captions in Pre-training

The impacts of long v.s. short captions on image-language pre-training. Training with short text-image pairs leaves certain tokens (e.g., garden token) easily overshadowed by salient tokens (e.g., castle token). long captions-image pairs can help bring the overshadowed tokens back into the light.

As the length of text (the number of sub-captions) increases, the performance of the pre-trained model on long-text-image retrieval consistently improves and becomes stable. However, there is degradation in the retrieval task and classification when the model is trained with longer text.

Pipeline

In order to find a solution that well balances both long and short texts, we design to add extra text tokens for text encoders, termed corner tokens ([Cor 1], [Cor 2], . · · · ), to aggregate diverse text features.


Experiments

Results on long-text-image retrieval tasks

BibTeX

@inproceedings{LoTLIP2024,
        title   = {LoTLIP: Improving Language-Image Pre-training for Long Text Understanding},
        author  = {Wu, Wei and Zheng, Kecheng and Ma, Shuailei and Lu, Fan and Guo, Yuxin and Zhang, Yifei and Chen, Wei and Guo, Qingpei and Shen, Yujun and Zheng-Jun, Zha},
        journal = {arXiv},
        year    = {2024}
    }