Generative Knowledge Selection for Knowledge-Grounded Dialogues

Findings of the Association for Computational Linguistics: EACL 2023, pages 2077–2088 May 2-6, 2023 ©2023 Association for Computational Linguistics Generative Knowledge Selection for Knowledge-Grounded Dialogues Weiwei Sun, Pengjie Ren, Zhaochun Ren∗ Shandong University, Qingdao, China sunnweiwei@gmail.com, {renpengjie,zhaochun.ren}@sdu.edu.cn Abstract Knowledge selection is the key in knowledge- grounded dialogues (KGD), which aims to se- lect an appropriate knowledge snippet to be used in the utterance based on dialogue his- tory. Previous studies mainly employ the clas- sification approach to classify each candidate snippet as “relevant” or “irrelevant” indepen- dently. However, such approaches neglect the interactions between snippets, leading to difficulties in inferring the meaning of snip- pets. Moreover, they lack modeling of the dis- course structure of dialogue-knowledge interac- tions. We propose a simple yet effective gener- ative approach for knowledge selection, called GENKS. GENKS learns to select snippets by generating their identifiers with a sequence-to- sequence model. GENKS therefore captures intra-knowledge interaction inherently through attention mechanisms. Meanwhile, we devise a hyperlink mechanism to model the dialogue- knowledge interactions explicitly. We conduct experiments on three benchmark datasets, and verify GENKS achieves the best results on both knowledge selection and response generation. 1 Introduction To improve the informativeness in open-domain dialogue agents (Freitas et al., 2020), knowledge- grounded dialogues (KGD) are proposed to lever- age external structured (Liu et al., 2019) and un- structured (Dinan et al., 2019) knowledge to di- alogue responses. In KGD, it is pivotal to em- bed factual and conversationally appropriate knowl- edge in responses. Two classes of approaches are considered to embed knowledge: end-to-end and pipeline. End-to-end models, such as FiD (Izac- ard and Grave, 2021), process the document and generate the response in one shot. However, they tend to misuse knowledge (Adolphs et al., 2021). Pipeline models address this problem by explic- itly identifying a specific knowledge snippet to ∗Corresponding author. Do you know Hedy Lamarr? Of course, she has been described as one of the great movie actresses! Which of her movies do you know? She has many classic films, such as Lady of the Tropics in 1939 and Boom Town in 1940. That is interesting! I'm going to watch these films! Besides being an actress, she also developed a radio guida nce system for Allied torpedoes a t the beginning of World War II. Hedy Lamarr (born Hedwig Eva Maria Kiesler; November 9, 1914 – January 19, 2000) was an Austrian-born American film actress and inventor. She has been described as one of the great movie actresses of all time. She became a film star with her performance in Algiers (1938). Her MGM films include Lady of the Tropics (1939), Boom Town (1940), H.M. Pulham, Esq. (1941), and White Cargo (1942). Her greatest success was as Delilah in Cecil B. DeMille's Bible-inspired Samson and Delilah (1949). ---------- She was honored with a star on the Hollywood Walk of Fame in 1960. ---------- At the beginning of World War II, she and composer George Antheil developed a radio guidance system for Allied torpedoes. This work led …

Content truncated to 500 words