ChatSearch: a Dataset and a Generative Retrieval Model for General Conversational Image Retrieval

TL;DR

A generative retrieval model named ChatSearcher is proposed, which is trained end-to-end to accept/produce interleaved image-text inputs/outputs and demonstrates strong capability in reasoning with multimodal context and can leverage world knowledge to yield visual retrieval results.

Abstract

In this paper, we investigate the task of general conversational image retrieval on open-domain images. The objective is to search for images based on interactive conversations between humans and computers. To advance this task, we curate a dataset called ChatSearch. This dataset includes a multi-round multimodal conversational context query for each target image, thereby requiring the retrieval system to find the accurate image from database. Simultaneously, we propose a generative retrieval model named ChatSearcher, which is trained end-to-end to accept/produce interleaved image-text inputs/outputs. ChatSearcher exhibits strong capability in reasoning with multimodal context and can leverage world knowledge to yield visual retrieval results. It demonstrates superior performance on the ChatSearch dataset and also achieves competitive results on other image retrieval tasks and visual conversation tasks. We anticipate that this work will inspire further research on interactive multimodal retrieval systems. Our dataset will be available at https://github.com/joez17/ChatSearch.

Figures (10)

• Figure 1: Our generative retrieval model ChatSearcher can accept multimodal inputs and generate textual response or retrieved images. ChatSearcher can comprehend multimodal dialogue context, infer user's implicit intentions, generate visual or textual responses through multimodal reasoning and world knowledge, and can also support interactive refinement of results.
• Figure 2: Illustration of automatic data construction pipeline for general conversational image retrieval dataset ChatSearch. We use foundation models (text generator GPT, image gallery retriever and image captioner) as generation tools to generate text dialogue and multimodal dialogue that aim at searching user desired images, as elaborated in \ref{['refgen']}. Then we apply context merging method and manually review on those data to construct a high quality evaluation split.
• Figure 3: Multimodal dialogue construction. The whole pipeline is combined with a text generator, a image search gallery and a pre-trained image captioner. The final output is a user-assistant multimodal conversation designed for image retrieval.
• Figure 4: Architecture of our generative retrieval model ChatSearcher. Interleaved documents serve as input, predicting words or retrieving images with generative training objective. Special token $\mathtt{[IMG]}$ predicts where to retrieve images. We use a dynamicaly-updated feature queue to save contrastive samples for image retrieval.
• Figure 5: Ablation study on feature queue size. We show the average recall on three retrieval benchmarks: Flickr30k, MSCOCO and our ChatSearch.