Sentence Punctuation for Collaborative Commentary Generation in Esports Live-Streaming

Hong Huang, Junjie Xu, Xiaoling Ling, Pujana Paliyawan

Abstract

To solve the existing sentence punctuation problem for collaborative commentary generation in Esports live-streaming, this paper presents two strategies for sentence punctuation for text sequences of game commentary, that is, punctuating sentences by two or three text sequence(s) originally punctuated by Youtube to obtain a complete sentence of commentary. We conducted comparative experiments utilizing and fine-tuning a state-of-the-art pre-trained generative language model among two strategies and the baseline to generate collaborative commentary. Both objective evaluations by automatic metrics and subjective analyses showed that our strategy of punctuating sentences by two text sequences outperformed the baseline.

Introduction

In the fast-paced world of Esports live-streaming, "collaborative commentary"—where an AI acts as a co-caster alongside a human—remains a significant challenge. A primary hurdle is the "fragmentation problem." Automated captioning systems, such as YouTube's, often slice live audio into short, incomplete text sequences based on arbitrary pauses rather than semantic meaning. Training AI models on these fragmented "Case Solo" snippets leads to nonsensical or ambiguous outputs, preventing the AI from functioning as a coherent partner in a live broadcast. This research explores re-grouping these snippets into more complete semantic units to improve the quality of generated commentary.

Methodology

The researchers investigated three data-grouping strategies to determine the optimal input length for a generative model:

  1. Case Solo (Baseline): Text sequences exactly as punctuated by YouTube.
  2. Case Duo: Grouping two consecutive YouTube sequences into a single training unit.
  3. Case Tri: Grouping three consecutive sequences into a single unit.

The study utilized the T5 (Text-to-Text Transfer Transformer) model, fine-tuning both "small" and "base" versions on a custom dataset of 100 League of Legends YouTube videos (90 for training, 10 for testing). The goal was to enable the model to predict and generate the next logical piece of commentary based on these punctuated inputs.

Comparison of generated commentary across Solo, Duo, and Tri punctuation strategies.

Key Insights and Results

The core finding is that the "completeness" of the input directly dictates the quality of the AI's output.

The "Goldilocks" Effect of Context While Case Solo often resulted in ambiguous phrases (e.g., "really frightens me but at the same time"), Case Duo allowed the model to generate specific, game-relevant phrases like "their strengths of being able to burst down the bear." However, Case Tri sometimes led to "over-generation," where the AI produced multiple commentaries at once, which could overwhelm a human co-caster.

Objective Performance Quantitative metrics (BLEU, ROUGE, and METEOR) confirmed the superiority of the Duo strategy. For instance, the fine-tuned T5-base model using Case Duo achieved a ROUGE-1 score of 16.60, a massive improvement over the 3.91 score of the Solo baseline. Fine-tuning was also shown to be essential for capturing specialized Esports jargon and player-specific actions.

Conclusion and Future Directions

The "Case Duo" strategy effectively mitigates the fragmentation problem, producing more human-like and contextually accurate commentary. Despite these gains, challenges remain regarding "rambling" in longer sequences (Case Tri) and the reliance on fixed-count heuristics for punctuation. Future research will focus on "intelligent" punctuation that identifies semantic boundaries and incorporating professional human evaluations to further refine the collaborative experience.

Summary updated

Abstract

To solve the existing sentence punctuation problem for collaborative commentary generation in Esports live-streaming, this paper presents two strategies for sentence punctuation for text sequences of game commentary, that is, punctuating sentences by two or three text sequence(s) originally punctuated by Youtube to obtain a complete sentence of commentary. We conducted comparative experiments utilizing and fine-tuning a state-of-the-art pre-trained generative language model among two strategies and the baseline to generate collaborative commentary. Both objective evaluations by automatic metrics and subjective analyses showed that our strategy of punctuating sentences by two text sequences outperformed the baseline.
BibTeX
@misc{huang2021sentence,
  author = {Hong Huang and Junjie Xu and Xiaoling Ling and Pujana Paliyawan},
  title = {{Sentence Punctuation for Collaborative Commentary Generation in Esports Live-Streaming}},
  year = {2021},
  eprint = {2110.12416},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2110.12416},
}
Figure