MatchTime: Towards Automatic Soccer Game Commentary Generation

TL;DR

This paper proposes a multi-modal temporal alignment pipeline to automatically correct and filter the existing dataset at scale, creating a higher-quality soccer game commentary dataset for training, denoted as *MatchTime* and trains an automatic commentary generation model, named **MatchVoice**.

Abstract

Soccer is a globally popular sport with a vast audience, in this paper, we consider constructing an automatic soccer game commentary model to improve the audiences' viewing experience. In general, we make the following contributions: First, observing the prevalent video-text misalignment in existing datasets, we manually annotate timestamps for 49 matches, establishing a more robust benchmark for soccer game commentary generation, termed as SN-Caption-test-align; Second, we propose a multi-modal temporal alignment pipeline to automatically correct and filter the existing dataset at scale, creating a higher-quality soccer game commentary dataset for training, denoted as MatchTime; Third, based on our curated dataset, we train an automatic commentary generation model, named MatchVoice. Extensive experiments and ablation studies have demonstrated the effectiveness of our alignment pipeline, and training model on the curated dataset achieves state-of-the-art performance for commentary generation, showcasing that better alignment can lead to significant performance improvements in downstream tasks.

Figures (7)

• Figure 1: Overview. (a) Left: Existing soccer game commentary datasets contain significant misalignment between visual content and textual commentaries. We aim to align them to curate a better soccer game commentary benchmark. (b) Right: While evaluating on manually aligned videos, existing models can achieve better commentary quality in a zero-shot manner. (The temporal window size is set to 10 seconds here.)
• Figure 2: Distribution of temporal offsets in our manually corrected SN-Caption-test-align. Through manual annotation, we find that the temporal discrepancy between the textual commentary and the visual content in the existing benchmark can even exceed 100 seconds.
• Figure 3: Temporal Alignment Pipeline. (a) Pre-processing with ASR and LLMs: We use WhisperX to extract narration texts and corresponding timestamps from the audio, and leverage LLaMA-3 to summarize these into a series of timestamped events, for data pre-processing. (b) Fine-grained Temporal Alignment: We additionally train a multi-modal temporal alignment model on manually aligned data, which further aligns textual commentaries to their best-matching video frames at a fine-grained level.
• Figure 4: MatchVoice Architecture Overview. Our proposed MatchVoice model leverages a pretrained visual encoder to encode video frames into visual features. A learnable temporal aggregator aggregates the temporal information among these features. The temporally aggregated features are then projected into prefix tokens of LLM via a trainable MLP projection layer, to generate the corresponding textual commentary.
• Figure 5: Qualitative results on commentary generation. Our MatchVoice demonstrates advantages in multiple aspects: (a) richer semantic descriptions, (b) full commentaries of multiple incidents in a single video, (c) accuracy of descriptions, and (d) predictions of incoming events.