Discussion Ontology: Knowledge Discovery from Human Activities in Meetings

Hironori TOMOBE
21st Century COE Program on Intelligent Media Integration, Nagoya University
Katashi NAGAO
Center for Information Media Studies, Nagoya University


Discussion mining is a preliminary study on gathering knowledge based on the content of face-to-face discussion meetings. To extract knowledge from discussion content, we have to analyze not only the surface arguments, but also semantic information such as a statement's intention and the discussion flow during meetings. We require a discussion ontology for this information. This discussion ontology forms the basis of our discussion methodology and requires semantic relations between elements in meetings. We must clarify these semantic relations to build the discussion ontology. We therefore generated discussion content and analyzed meeting metadata to build the ontology.

1 Introduction

Meetings are a normal part of business and academic settings.For example, companies have meetings, which involve deciding the time and place beforehand. However, they also have casual meetings, which neither the time nor place is decided.The themes of meetings are not only for decision making but also for brainstorming to have participants generate ideas.Meetings are indispensable as a place for exchanging opinions with others and for forming a social network.

Knowledge must be shared efficiently in brainstorming meetings. Meetings in which participants obtain a lot of knowledge over a short period of time tend to be evaluated highly, while meetings in which participants obtain only a little knowledge even though they spend a long time in them tend not to be evaluated highly.Therefore we need to determine the elements that are necessary to hold efficient meetings.

We hypothesize that three elements are important for holding efficient meetings: specifying information for on-going discussion, generating minutes that have high reproduction value, and improving the skill of the participants.To achieve these three things, we require a discussion ontology, and this ontology requires that we clarify the semantic relations between elements in a meeting.

To build a discussion ontology, we generated discussion contents with structure for a meeting and analyzed the metadata in the contents. Using browser-based tools for acquiring information about meetings in the real world, we structurized the discussion contents. We inferred the intentions of statements in a meeting based on classification rules generated by a machine learning algorithm and extracted the discussion flows.

2 Discussion Ontology

We define the description of discussion elements and semantic relations between these elements as a discussion ontology. Each meeting is composed of discussions, and each discussion is composed of statements. We define these discussions and statements as discussion elements. Similarly, participants and meeting materials (such as presentation slides and handouts) are defined as discussion elements. Semantic relations exist between these elements: for example, dependence exists between discussions and statements, similar relations exist between statements, and correspondence relations exist between meeting materials and discussions.

In addition, the semantic relations of discussion elements include some restrictions. For example, restrictions such as ``After a participant asks a question (one intention), another one must answer (another intention)'' are important for keeping the discussion flowing smoothly in meetings. A system can determine the on-going discussion situation and assist with meetings by specifying this situation by clarifying the semantic relations and restrictions. The discussions are also formalized. We attempted to segment discussion based on discussion elements and to clarify semantic relations and restrictions on discussion elements by accumulating information about meetings and by analyzing these metadata.

3 Gathering Metadata for Generating Discussion Content

3.1 Structuring Discussions

Meetings in the real world (i.e., face-to-face meetings) progress on the basis of a time series. Discussion elements like statements and discussions, however, cannot be expressed in a time series. Therefore, discussions need to be properly structured to understand discussions semantically.

Structurization of discussions means clarifying elements that make up a discussion and the semantic relations between the elements. In other words, structurization of discussions involves segmenting meetings based on discussion elements and adding links between segments. The unit of segmentation for discussions and the semantic relations of each element are different for each meeting. Therefore, this study deals with meetings that have presentations because, in these meetings, the moderator projects her/his presentation material on a screen, making the targets of discussion clearer.

However, extracting discussion elements and linking the elements is difficult. Therefore, we used discussion mining to structure discussions semi-automatically.

3.2 Discussion Mining

Most studies on providing technology for discussions and generating minutes have focused on automatic recognition techniques for audio and visual data, such as meeting browsers. We used a discussion mining system for generating discussion content. Figure shows an image of the discussion room.

Discussion Mining System

Fugure1: Discussion Mining System

We targeted face-to-face meetings in the discussion mining. Detailed recording is done during the meeting with four cameras and a microphone that are installed in the meeting room. One camera records the main screen, another camera records the presenter's face, and the two remaining cameras record the participants' actions. Audio information is recorded using a microphone installed in the center of the meeting room. Participants in the meeting transmit their IDs and comments using tag devices called discussion tags to enable the discussion to be properly structured. Furthermore, using a button device, the participants can arbitrarily input their stance toward the presentation and any arguments with other participants. This information is added to the minutes in real time and is edited by the secretary. Currently, the secretary inputs text manually, though in the future, this will be done automatically with speech recognition technology. A record of the arguments in XML and MPEG-4 format is saved in an XML database as multimedia minutes.

here are two types of discussion tags: ``start-up'' and ``follow-up.'' Participants use the startup tag to make remarks that trigger a discussion and use the follow-up tag to make remarks that relate to an ongoing discussion. The system uses these tags to segment the discussion automatically to enable an analysis of the discussion and viewing of the video.

3.3 Usage of Discussion Content

his system enables us to visualize the structure of minutes by creating a graphical display and edit mode for statements with the use of scalable vector graphics (SVG). The graph is semi-automatically structured with pertinent information and keywords of statements and slides, as shown in Fig. . This function allows users to edit the content.

Graph Viewer of Discussion Content

Fugure2: Graph Viewer of Discussion Content

The detailed discussion can be understood from watching the video recorded by the discussion mining system. However, efficiently inspecting the video is a problem. Our discussion mining system provides a tool for effectively viewing discussion content recorded in the meeting. This system provides a meeting video with a support system for discussion content browsing (shown in Fig. ).

Support System for Discussion Content Browsing

Fugure3: Support System for Discussion Content Browsing

4 Classification of Statements and Extraction of Discussion Flow

Each statement in a discussion consists of two kinds of information: the topic and the intention. The former is important for understanding the discussion theme and specific content. The latter is important for understanding the discussion flow. In this section, we describe a method for extracting the intention. Knowing the intention of statements should enable understanding the discussion flow. Using this flow and adding the importance of discussion will enable users to find important discussions easily.

The procedure of this method is as follows. First, we determine the intention of each statement. We define an intention tag set for meetings from the information on meetings that we have accumulated over the previous four years. Each intention tag expresses the intention included in each statement, and we limit the number of types of intention tags. Second, we classify each statement based on its intention tag. We infer intention tags for statements by using attributes of the statements such as speaker and type of statement, because adding intention tags to all statements manually is time consuming and difficult. Finally, after we add intention tags to all statements, we find the intention tags that often co-occur in a discussion.

4.1 Intention Tag

The discussion can be analyzed semantically by adding intention tags to each statement in a meeting. We can use intention tags to understand the discussion flow, and the features of face-to-face communication can be clarified by analyzing the local discussion structure, such as ``requests'' and ``replies.''

The intention tag set needs to be based on a standard tag set to improve reliability. The tag set naturally varies based on the usage, but we must add tags without any ambiguity. We created intention tag sets based on a standard utterance-unit tagging scheme .

We manually added standard utterance-unit tags to all four-minute statements, and we define the tag set to use based on the frequency distribution of the standard utterance-unit tag.

  • Request for unknown information

    a statement in which the speaker wants some values as a reply

  • Request for true or false information

    a statement in which the speaker wants ``yes'' or ``no'' as a reply

  • Reply to unknown information

    a statement in which the speaker replies to a request for unknown information

  • a statement in which the speaker replies to a request for true or false information

    Reply to true or false information

  • a statement in which the speaker does not reply directly to the request for information

    Reply suspension

  • Information presentation

    a statement in which the speaker presents her/his knowledge

  • Suggestion/Proposal

    a statement in which the speaker suggests or proposes actions to be taken by other participants

  • Opinion/Hope

    a statement that presents the speaker's ideas or desired outcome

4.2 Classification of Statements

We stored more than 370 minutes of meetings in the discussion mining system, and about 14,500 statements are included in the minutes. Because classifying these statements into an intention tag set manually is difficult, we inferred which tags were added to each statement by using a machine learning algorithm, C4.5 .

We used the following metadata of statements as attributes: ``speaker (presenter or participants except for the presenter),'' ``type of statement (start-up or follow-up),'' ``speaking time,'' ``occurrence of keywords,'' etc. In addition, we added an intention tag to three minutes manually to use them as training data and generated classification rules for the intention tag. Examples of the classification rules are shown below.

  • type of statement = start-up, speaker = participants except for the presenter, occurrence of keywords = yes, speaking time = short (less than 20 seconds

    Request for unknown information

  • Opinion/Hope

    ype of statement = start-up,speaker = participants except for the presenter,occurrence of keywords = no,speaking time = mid-length (21 - 40 seconds)

When a statement had more than two types of intention tags as a result of applying the classification rules, we added all of the intention tags to the statement because there are many statements that should have several intention tags. For example, after one participant speaks with the intention ``Request for unknown information,'' another participant may reply with an intention that includes not only the ``Reply to unknown information'' but also an ``opinion/hope.'' The classification rules' precision was 83%; we tuned the intention tags manually after applying the classification rules.

4.3 Extraction of Discussion Flow Based on Statement Intention

The statement type, either start-up or follow-up, is added to each statement using discussion tags in the discussion mining system. All discussions must start with a start-up statement and continue with a follow-up statement when participants discuss topics continuously. We define a discussion that has one start-up statement and related follow-up statements as a unit of discussion, and we call this a discussion segment.

We extracted discussion flows from 20 minutes and investigated the flows that appeared frequently and the co-occurrences of statement intention. Table shows the ten flows that appeared frequently.

e believe that a discussion flow that appears frequently is a model of efficient discussion in meetings. This kind of discussion flow is part of the grammar of the discussion, and the discussion according to this grammar will obtain better results for participants in meetings. In contrast, a discussion flow that deviates from this grammar could become a diverging discussion, thereby preventing the meeting from being efficient.

Discussion flows have several potential applications. They can be used to detect the similarities and differences between discussions. A discussion flow can be considered a feature of a discussion, and a discussion grammar can be defined by using this feature. The discussion grammar can be used to summarize the discussion and prepare the minutes. For example, when a discussion in which questions and answers are repeated has the flow ``answer - question - answer - question -...,'' it can be summarized by omitting this repetition pattern.

Discussion flows have several potential applications. They can be used to detect the similarities and differences between discussions. A discussion flow can be considered a feature of a discussion, and a discussion grammar can be defined by using this feature. The discussion grammar can be used to summarize the discussion and prepare the minutes. For example, when a discussion in which questions and answers are repeated has the flow ``answer - question - answer - question -...,'' it can be summarized by omitting this repetition pattern.

5 Conclusion and Future Work

We have described a method for classifying statements in discussions and extracting discussion flows. Analysis of the metadata of a structured discussion clarified the importance of the discussion elements and revealed information necessary for understanding discussions. Moreover, the semantic relations between discussion elements, which consist of discussion content, can be clarified, and a discussion ontology can be built. We believe that natural language processing can be used to discover knowledge efficiently based on the topics of discussions.

The minutes of a meeting and any audio and visual data recorded in the meeting play an important role because from them we can record the details of the discussions. We therefore plan to clarify the role of audio and visual data in discussion ontology.