CAMEO: Composed Access to Multimodal E-commerce Objects

This is a pilot task on Composed Access to Multimodal E-commerce Objects (CAMEO) at the NTCIR-19 Conference, focusing on enhancing e-commerce search and question answering through innovative multimodal approaches.

Dataset Exploration Tool

Explore the ViEcom-Rec dataset. This tool allows you to interactively explore the dataset, view product details, reviews, and other relevant information to better understand the data you'll be working with.

Overview

CAMEO presents two exciting subtasks, each designed to test your innovative solutions:

1. Composed Image Retrieval (CIR)

Gone are the days of simple image search! In CIR, users provide a reference product and a textual modification to find a altered variant. Think of it as empowering users to describe exactly what they want, beyond just what they see.

  • Your Mission: Given a product ID and a textual modification (e.g., "same type, but a larger size"), your system must return a ranked list of products that precisely match the revised intent.
  • Example Query: "I want the same product, but made in Vietnam"
  • Evaluation: Your performance will be measured using industry-standard metrics like Precision, Recall, and mean Average Precision (mAP) at K.

2. Review-Based Question Answering (QA)

Online reviews are a treasure trove of information, but extracting specific answers can be a challenge. This subtask focuses on helping users get direct answers to their product questions from customer reviews.

  • Your Mission: Given a product ID or image and a user question (e.g., "Is this suitable for dry skin?"), your system needs to provide a two-part answer:
    1. A concise natural language answer (e.g., "Yes, it is suitable for dry skin.").
    2. A list of supporting evidence sentences extracted from one or more product reviews.
  • The Unique Twist: Unlike many existing datasets, CAMEO encourages you to aggregate evidence across multiple user reviews, reflecting how real consumers form opinions from diverse feedback!
  • Evaluation: Your short answer will be evaluated using Exact Match (EM) and F1 score, while your evidence list will be assessed using sentence-level precision, recall, and F1.

The ViEcom-Rec Dataset

Both subtasks are powered by the ViEcom-Rec dataset, a robust collection of real-world e-commerce data from a Vietnamese platform.

  • Dataset Scale: Over 369,000 Vietnamese-language product reviews from 304,708 users for 2,244 distinct products.
  • Rich Product Information: Includes product images, reviews, seller info, product descriptions, and nine key attributes extracted by GPT-family models (item name, ingredient, product feature, skin type, capacity, design, brand, expiry, and origin).
  • Expanding Diversity: While initially focused on face cleansers, we are expanding the dataset to include a wider range of cosmetics categories like sun cream, lipsticks, and face washers to better match real-world scenarios!
  • Open Access: The ViEcom-Rec dataset is publicly available under a Creative Commons Attribution-NonCommercial-ShareAlike 4.0 license for non-commercial academic use at Github.
  • Bridging the Language Gap: To ensure maximum participation, all Vietnamese text data will be accompanied by high-quality English translations, and task guidelines, evaluation scripts, and starter code will be provided in English