Diverse and Faithful Knowledge-Grounded Dialogue Generation via Sequential Posterior Inference
Jul 24, 2023
Sprecher:innen
Deqian Kong
Sprecher:in · 0 Follower:innen
Pascale Fung
Sprecher:in · 0 Follower:innen
Ying Nian Wu
Sprecher:in · 0 Follower:innen
Über
The capability to generate responses with diversity and faithfulness using external knowledge is of vital importance in building a human-like trustworthy dialogue system. One line of the popular approaches is to adopt a two-step paradigm by optimizing knowledge selection and response generation separately, which may neglect the inherent correlation between these two tasks. Another line of research leverages a conditional variational model to optimize knowledge selection and response generation jointly by recruiting an inference network.In this paper, we propose an end-to-end learning framework, Sequential Posterior Inference (SPI), that is able to select knowledge and generate dialogues by approximately sampling from the posterior distribution. It does not require the inference network or assume a simple geometry of the posterior distribution. This simple and natural inference procedure of SPI by directly querying the response generator is able to make accurate knowledge selection and generate fruitful responses. We further equip SPI with an initializer and short-run inference dynamics to explore the discrete and continuous search space efficiently and effectively. Besides theoretical niceties, experimental results on two common dialogue datasets (Wizard of Wikipedia and Holl-E) show that SPI is superior to previous strong methods on both automatic and human evaluation metrics.The capability to generate responses with diversity and faithfulness using external knowledge is of vital importance in building a human-like trustworthy dialogue system. One line of the popular approaches is to adopt a two-step paradigm by optimizing knowledge selection and response generation separately, which may neglect the inherent correlation between these two tasks. Another line of research leverages a conditional variational model to optimize knowledge selection and response generation j…
Gefällt euch das Format? Vertraut auf SlidesLive, um euer nächstes Event festzuhalten!
Professionelle Aufzeichnung und Livestreaming – weltweit.
Freigeben
Empfohlene Videos
Präsentationen, deren Thema, Kategorie oder Sprecher:in ähnlich sind
Bit Allocation using Optimization
Später ansehen
Tongda Xu, …
A New PHO-rmula for Improved Performance of Semi-Structured Networks
Später ansehen
NP-SemiSeg: When Neural Processes meet Semi-Supervised Semantic Segmentation
Später ansehen
Quantile Credit Assignment
Später ansehen
Scalable Safe Policy Improvement via Monte Carlo Tree Search
Später ansehen
