SPORTSINTERVIEW: A Large-Scale Sports Interview Benchmark for Entity-centric Dialogues

Proceedings of the 13th Conference on Language Resources and Evaluation (LREC 2022), pages 5821–5828 Marseille, 20-25 June 2022 © European Language Resources Association (ELRA), licensed under CC-BY-NC-4.0 5821 SPORTSINTERVIEW A Large-Scale Sports Interview Benchmark for Entity-centric Dialogues Hanfei Sun, Ziyuan Cao, Diyi Yang Georgia Institute of Technology Atlanta, GA hsun315@gatech.edu, zcao300@gatech.edu, diyi.yang@cc.gatech.edu Abstract We propose a novel knowledge grounded dialogue (interview) dataset SPORTSINTERVIEW set in the domain of sports interview. Our dataset contains two types of external knowledge sources as knowledge grounding, and is rich in content, containing about 150K interview sessions and 34K distinct interviewees. Compared to existing knowledge grounded dialogue datasets, our interview dataset is larger in size, comprises natural dialogues revolving around real-world sports matches, and have more than one dimension of external knowledge linking. We performed several experiments on SPORTSINTERVIEW and found that models such as BART fine-tuned on our dataset are able to learn lots of relevant domain knowledge and generate meaningful sentences (questions or responses). However, their performance is still far from humans (by comparing to gold sentences in the dataset) and hence encourages future research utilizing SPORTSINTERVIEW. Keywords: Entity-centric sports interview, conversation dataset, text generation, deep neural network 1. Introduction Recent years, large-scale corpora containing human- human interaction made the modeling of human dialogues possible (Li et al., 2017). To think about it, as humans, most of our daily dialogues are grounded on external knowledge we are aware of. Hence, it’s natural for previous works to focuses on incorporating external knowledge in the task of response generation (Dinan et al., 2018; Zhou et al., 2018). However, most of these datasets only contain external knowledge from one knowledge source (one dimension), which limits the diversity of knowledge sources and may contain unwanted bias. On the other hand, we humans can eas- ily acquire knowledge from multiple sources, like from Internet search or from others’ words in a conversation. Putting this concern for a while, there are actu- ally many different forms of dialogue, ranging from chatting on the Internet to group discussions. In partic- ular, interview dialogues are a specific kind of natural dialogue consisting of interviewers and interviewees chatting structurally about certain topics (usually focusing on recently occurring events). It differs from open-domain chit-chat in that interviewers and inter- viewees have different responSportsInterviewilities during the conversation. Specifically, the interviewer typically brings up a topic first, which the interviewee will expand on and talk about his/her perspectives about it. Then the interviewer can choose to follow up on that topic or change the topic altogether to guide the flow of the conversation. Interview dialogues are great candidates for knowledge-grounded dialogue modeling tasks as the discussions usually resolve around some certain topic and often go much deeper, requiring sufficient domain knowledge. There have been recent efforts to collect a large-scale interview QUESTION: Just a high ankle sprain just lingered and lingered? SAMAJE PERINE: Yeah. QUESTION: Is that one of those things that you deal with a lot? I know in high school you carried the …

Content truncated to 500 words