Design And Implementation Of Automatic Comment Generation System Towards Social Media

Recent years have witnessed the flourish of social media,like Twitter,Facebook,and Instagram,where governments,organizations and ordinary people can generate social posts and comments to manage crises,promote products and build up relationships,respectively.In this paper,we argue that automatic comment generation for posts in social media is in urgent need because of the following reasons:1)facilitating governments to guide the public opinion regarding specific social events by automatically generating comments and thus maintain the social stability;2)saving great efforts for companies to brand their newly released products by injecting their advertisements into the generated comments;and 3)providing references to ordinary people,who do not know how to express appropriately for the time being,but really expect to interact with the user of the given tweet.Automatic comment generation is,however,non-trivial due to the following facts.Firstly,different people tend to have different linguistic styles.For example,in the context of social media,some people are keen to use imperative sentence with the consecutive punctuation marks,and some emojis.Therefore,how to model and simulate users' linguistic styles for personalized comment generation poses a main challenge for us.Secondly,in practice,people,more often than not,reply a given tweet with certain emotion,namely explicitly negative or positive polarities.Inspired by this,how to enable the generated comments with designated emotion polarity constitutes another research challenge.Last but not least,there is no publicly available large-scale dataset to well support the validation of the personalized and emotionalized comment generation.To tackle the aforementioned challenges,we present a novel scheme of emotion-aware personalized comment generation for social media posts,dubbed as CRobot.In particular,CRobot consists of two pivotal components emotion-aware user linguistic style modeling and multi-adversarial personalized comment generation.To be more specific,the former works on learning the latent representations of the emotion-aware historical tweets and the avatar of the user,which targets at characterizing capture the user's emotion-aware linguistic style.Whereas,the latter is devised to generate proper personalized comments by jointly regularizing multi-adversarial discriminators,namely the general comment discriminator and emotion-aware personalized discriminator.It is worth noting that we subtly embed the adversarial learning within the reinforcement learning framework to boost the generation performance.In this context,CRobot aims to tackle the comment generation problem by taking a sequence of actions,where a word is selected from the token vocabulary based on the current state,i.e.,the previous output tokens,the given user,and the given emotion polarity,at each time step.To thoroughly justify our proposed CRobot model,we have constructed a large-scale and real-world dataset based upon Twitter,comprising 6,763 tweets with 1,461,713 corresponding comments created by 153,664 users.Extensive experiments on this real-world dataset have demonstrated the superiority of our model over several state-of-the-art baselines.