Developing a POI Recommendation Skill for a Conversational Conference Assistant

Abstract

In this project we developed a skill for recommending points of interest through conversations with users, suitable for a digital assistant stationed at a public venue. The skill was part of a larger effort involving several teams working on a conversational conference assistant called DAGFiNN, which was exhibited at the 2022 European Conference on Information Retrieval (ECIR). The assistant was developed on the open-source conversational AI platform called Rasa and implemented on a social robot made by Furhat Robotics alongside an external display. Our skill focuses on sequential recommendations through slot-filling, and allows for related follow-up questions along the way, such as asking for transport or walking distance to the recommended POI. Rich responses shown on the display includes subtitles for the dialogue, locations on a map, QR-codes for links, and other useful information. The skill was also made to be easily re-adaptable for other public events, by changing only the database and a few values in the source code.

Results suggested that the skill was successful in providing effective and accurate recommendations based the user's preference, but that lack of system-initiated guidance could leave the user confused about its scope.

In this project we developed a skill for recommending points of interest through conversations with users, suitable for a digital assistant stationed at a public venue. The skill was part of a larger effort involving several teams working on a conversational conference assistant called DAGFiNN, which was exhibited at the 2022 European Conference on Information Retrieval (ECIR). The assistant was developed on the open-source conversational AI platform called Rasa and implemented on a social robot made by Furhat Robotics alongside an external display. Our skill focuses on sequential recommendations through slot-filling, and allows for related follow-up questions along the way, such as asking for transport or walking distance to the recommended POI. Rich responses shown on the display includes subtitles for the dialogue, locations on a map, QR-codes for links, and other useful information. The skill was also made to be easily re-adaptable for other public events, by changing only the database and a few values in the source code.

Results suggested that the skill was successful in providing effective and accurate recommendations based the user's preference, but that lack of system-initiated guidance could leave the user confused about its scope.