Generating Hashtags for Short-form Videos with Guided Signals

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 9482–9495 July 9-14, 2023 ©2023 Association for Computational Linguistics Generating Hashtags for Short-form Videos with Guided Signals Tiezheng Yu†∗, Hanchao Yu‡, Davis Liang‡, Yuning Mao‡, Shaoliang Nie‡, Po-Yao Huang‡, Madian Khabsa‡, Pascale Fung†, Yi-Chia Wang‡ †Hong Kong University of Science and Technology ‡Meta AI tyuah@connect.ust.hk, yichiaw@meta.com Abstract Short-form video hashtag recommendation (SVHR) aims to recommend hashtags to con- tent creators from videos and corresponding descriptions. Most prior studies regard SVHR as a classification or ranking problem and se- lect hashtags from a set of limited candidates. However, in reality, users can create new hash- tags, and trending hashtags change rapidly over time on social media. Both of these properties cannot be easily modeled with classification approaches. To bridge this gap, we formulate SVHR as a generation task that better repre- sents how hashtags are created naturally. Ad- ditionally, we propose the Guided Generative Model (GGM) where we augment the input features by retrieving relevant hashtags from a large-scale hashtag pool as extra guidance sig- nals. Experimental results on two short-form video datasets show that our generative mod- els outperform strong classification baselines, and the guidance signals further boost the per- formance by 8.11 and 2.17 absolute ROUGE- 1 scores on average, respectively. We also perform extensive analyses including human evaluation, demonstrating that our generative model can create meaningful and relevant novel hashtags while achieving state-of-the-art per- formance on known hashtags 1. 1 Introduction Short-form videos on social media are increasingly popular thanks to the proliferation of multimedia technologies and portable devices (Vandersmissen et al., 2014; Montag et al., 2021). To highlight the topics and salient information of the videos, hash- tags – words or unspaced phrases prefixed with a “#” – have been widely used. Proper use of hashtags can also increase the probability of the videos being discovered (Cao et al., 2020). In light of this, short-form video hashtag recommendation ∗ Work is done during an internship at Meta. 1The code is released at: https://github.com/ facebookresearch/hashtag-generation Video Frames: GGM Generation Results: #snowstorm #winter #cold-weather #toronto Video Description: First Snow Storm of 2008 in Toronto, Ontario, Canada. Ground Truth Hashtags: #canada #snowstorm #toronto #winter V-MLM Classification Results: #snow #winter #snowstorm #canada #2008 Figure 1: Video frames and the video description are the inputs. The Guided Generative Model (GGM) gen- erates hashtags related to video frames (e.g., #winter and #cold-weather) as well as video description (e.g., #snowstorm and #toronto). The generated novel hashtag that never appears in the training set is highlighted. We use mock-up video frames in the paper. (SVHR), which aims to suggest relevant and mean- ingful hashtags to content creators when they share videos, has received considerable attention from industry and academia (Li et al., 2019; Jain and Jindal, 2020; Mehta et al., 2021). However, most previous studies on SVHR consider it as a classifi- cation problem and rank the hashtags in a small and fixed-size set one by one (Li …

Content truncated to 500 words