Not All Comments Are Equal: Insights into Comment Moderation from a Topic-Aware Model

Proceedings of Recent Advances in Natural Language Processing, pages 147–154 Sep 1–3, 2021. https://doi.org/10.26615/978-954-452-072-4_018 147 Neural Network-Based Generation of Sport Summaries: a Preliminary Study David St´ephane Belemkoabga*, Aur´elien Bossard**, Abdallah Essa*, Christophe Rodrigues*andK´evin Sylla* *L´eonard de Vinci Pˆole Universitaire, Research Center 92 916 Paris, La D´efense, France, firstname.name@devinci.fr **Laboratoire d’Informatique Avanc´ee de Saint-Denis, Universit´e Paris 8 2 rue de la Libert´e, 93526 Saint-Denis, France firstname.name@iut.univ-paris8.fr Abstract This paper presents a global summarization method for live sport commentaries for which we have a human-written summary available. This method is based on a neural generative summarizer. The amount of data available for training is limited compared to corpora com- monly used by neural summarizers. We pro- pose to help the summarizer to learn from a limited amount of data by limiting the entropy of the input texts. This step is performed by a classification into categories derived by a detailed analysis of the human-written sum- maries. We show that the filtering helps the summarization system to overcome the lack of resources. However, several improving points have emerged from this preliminary study, that we discuss and plan to implement in future work. 1 Introduction In this paper, we propose a new approach for sport commentary summarization. This approach is still an ongoing work, and the results presented here are preliminary. Sport commentaries represent an interesting resource, as the live commentaries we work on are associated with a summary written by an expert : the commentator himself. Indeed, the commentator writes a summary at the end of every game. Automatically generating a game summary would release the commentator from a part of his heavy workload and thus would free up his time for more complex and rewarding tasks, such as game in-depth analysis. Summarizing live sport events commentaries is a challenging task. First of all, they are live written; new commentaries can conflict with or complete former ones. For example, if a soccer player scores two goals, the live commentaries about each goal are relevant information. However, extracting the live commentaries about each goal will not be suffi- cient in order to generate a good summary. It would indeed lead to producing redundant summaries. Moreover, if you consider a game as an event, it is composed of several subevents. Some of them are deemed important enough to be commented. However, most of those commented subevents are not important enough to appear in a summary. So, live commentaries are mostly made up of noise : in a soccer game, there will be more shots than goals, even more substitutions than goals, which are the most important information of a game. This noise has to be filtered. Therefore, studies have to be carried out about the relevance of an information in the game summarization context. The style of human-written summaries differs from the one of live commentaries. For all these reasons, statistical extractive summarization mod- els are not relevant for this kind of data. Extractive models consist in fact in extracting relevant sen- tences from source …

Content truncated to 500 words