Training and Evaluating Multimodal Word Embeddings with Large-scale Web Annotated Images

TL;DR

This paper introduces a large-scale multimodal dataset combining text and images from Pinterest, and proposes RNN-based models that incorporate visual information to improve word embeddings, evaluated on semantic similarity tasks.

Contribution

The paper presents a new large-scale dataset with 300 million sentences and 40 million images, and develops multimodal RNN models that effectively integrate visual data into word embeddings.

Findings

Visual information improves word embedding quality

Weight sharing enhances multimodal learning

Large-scale dataset enables better semantic similarity evaluation

Abstract

In this paper, we focus on training and evaluating effective word embeddings with both text and visual information. More specifically, we introduce a large-scale dataset with 300 million sentences describing over 40 million images crawled and downloaded from publicly available Pins (i.e. an image with sentence descriptions uploaded by users) on Pinterest. This dataset is more than 200 times larger than MS COCO, the standard large-scale image dataset with sentence descriptions. In addition, we construct an evaluation dataset to directly assess the effectiveness of word embeddings in terms of finding semantically similar or related words and phrases. The word/phrase pairs in this evaluation dataset are collected from the click data with millions of users in an image search system, thus contain rich semantic relationships. Based on these datasets, we propose and compare several Recurrent Neural Networks (RNNs) based multimodal (text and image) models. Experiments show that our model benefits from incorporating the visual information into the word embeddings, and a weight sharing strategy is crucial for learning such multimodal embeddings. The project page is: http://www.stat.ucla.edu/~junhua.mao/multimodal_embedding.html