Meeting Event Recognition using a Parallel Recurrent Neural Net Approach

In this paper we present a novel architecture for the task of meeting event recognition in meetings that was inspired by the neural field theory. These group actions provide a basis that enables effective browsing and querying in a meeting archive. For our research we used the public available meeting corpus that is described in [3]. This corpus consists in special scripted meetings that were recorded in the IDIAP Smart Meeting Room. Each recorded meeting consists of a set of predefined group actions in a specific order that was defined in an agenda. The appearing group actions are: Monologue (one participant speaks continuously without interruption), Discussion (all participants engage in a discussion), Note-taking (all participants write notes), White-board (one participant at front of room talks and makes notes on the white board), and Presentation (one participant at front of room makes a presentation using the projector screen). A total of 53 scripted meetings with two disjoint sets of meeting participants were recorded. 30 of them were used for the training, the remaining 23 videos were used for the evaluation of the system. In each meeting there were four participants at six possible positions: four seats plus whiteboard and presentation board. For the task of segmenting the meetings into group actions we propose a new approach that is based on the theory of the neural fields, first analyzed by Amari [1]. The idea is to present the features of a whole meeting to the neural field simultaneously and get a segmentation and classification as output. In this way elements from the end of a meeting can have influence on elements at the beginning and the other way round which should increase the robustness of the classification task. The typical equation for a neural field is denoted in eq. 1.