The Multimodal Sentiment Analysis in Car Reviews (MuSe-CaR) Dataset: Collection, Insights and Improvements

TL;DR

This paper introduces MuSe-CaR, a comprehensive multimodal dataset for real-life car review sentiment analysis, providing insights into data collection, annotation, and a novel model that significantly improves trustworthiness prediction.

Contribution

The paper presents MuSe-CaR, the first large-scale multimodal dataset for in-the-wild sentiment analysis in car reviews, along with a new Multi-Head-Attention model for improved trustworthiness prediction.

Findings

The dataset covers audio-visual and language modalities in real-world scenarios.

The proposed Multi-Head-Attention model outperforms baseline by nearly 50%.

No participant outperformed the baseline in trustworthiness prediction, highlighting the challenge.

Abstract

Truly real-life data presents a strong, but exciting challenge for sentiment and emotion research. The high variety of possible `in-the-wild' properties makes large datasets such as these indispensable with respect to building robust machine learning models. A sufficient quantity of data covering a deep variety in the challenges of each modality to force the exploratory analysis of the interplay of all modalities has not yet been made available in this context. In this contribution, we present MuSe-CaR, a first of its kind multimodal dataset. The data is publicly available as it recently served as the testing bed for the 1st Multimodal Sentiment Analysis Challenge, and focused on the tasks of emotion, emotion-target engagement, and trustworthiness recognition by means of comprehensively integrating the audio-visual and language modalities. Furthermore, we give a thorough overview of the dataset in terms of collection and annotation, including annotation tiers not used in this year's MuSe 2020. In addition, for one of the sub-challenges - predicting the level of trustworthiness - no participant outperformed the baseline model, and so we propose a simple, but highly efficient Multi-Head-Attention network that exceeds using multimodal fusion the baseline by around 0.2 CCC (almost 50 % improvement).