This Goecities web site has been running since 1998, and has given stirling service. Times change, however, and I have outgrown this space. Please visit my shiny new site www.paulrudman.net which is under construction. Meanwhile, I am leaving this space as a snapshot of 2000-2002 when I lived in Birmingham and only hoped to make it to where I am now-with a smart "Dr." title. Ultimately, I will reuse this web space for a new purpose. Watch this space! Paul

"Automatic Support for Conversations using Semantic Relationships

The area being investigated is that of one adult learning from a more knowledgeable person (peer, tutor, expert etc.) In such a situation, the ‘knowledgeable person’ (KP) has ready access to information as yet unknown to the learner. The purpose of the conversation is to first identify which information needs transferring, and then help the learner to assimilate that information. This is often achieved in stages; each time new information closest to the limit of the learner’s existing knowledge is identified, transferred and assimilated. This expands the limit of the learner’s existing knowledge, allowing the next transfer of information be nearer to the target information, i.e. that which the learner seeks to learn. The process is negotiated between the KP and learner, whereby the KP elicits information about the learner’s existing knowledge, assesses the most useful piece of information to impart next, and provides that information. This is followed by feedback from the learner as to the understandability and relevance of the information. The process then repeats until either the learner reports that the target has been reached or one of the conversants ends the conversation.

During this process, the learner has the problem of not knowing what to ask, since this would require domain knowledge they have yet to learn. The KP, on the other hand, has the problem of not knowing the current extent of the learner’s knowledge, and thus which information to impart next. Further, the situation differs from face-to-face conversations in lacking visual cues about the other person’s mental state (interested / confused / thoughtful / confident, etc.) making the negotiation process more difficult.

The PhD described here investigates the utility of providing small amounts of relevant information to the learner during the negotiation stages of the conversation. The information is basically a list of domain areas semantically close to the area under discussion. Thus, the learner is made aware of the KP’s options and can participate more directly in the decision of which domain area the KP will explain next.

In order to achieve this using standard computer technologies, three problems need to be solved. Firstly, information about the learner’s current knowledge must be obtained. Secondly, domain areas semantically close to the area under discussion must be identified. Finally, summary information about these areas must be presented to the learner without disrupting the ongoing conversation.

1) Obtaining information about the learner’s current knowledge. Two method have been investigated. Firstly, the learner was asked to externalise their learning using a concept map. The creation of a node on the concept map was taken as evidence that learning had taken place in the domain area defined by the text. The second method was to use commercially available speech recognition to identify words used by the KP, as (s)he explained a domain area. This was taken as a pointer to the current conversation domain area (not necessarily that that learning had taken place). It is assumed that domain areas learned will be a subset of domain areas explained.

2) It is necessary for the computer to be able to link domain areas automatically. i.e., given one piece of information, there needs to be a definition of related information. This is achieved using a concept map of the domain. A complete concept map is created in computer-readable format. (JAVA-based software was purpose-written to control all aspects of the experimental work described here). During a conversation, the learner creates a node on his/her (initially blank) concept map. The computer then compares the learner’s newly created node against those on the computer’s map (which is hidden from both learner and KP). Identifying a match implies that the learner is currently working in that domain area. In the second case, where speech recognition is used to monitor the KP’s speech, each new word spoken is used, along with other recently spoken words, to dynamically assess which node on the computer’s map is currently being discussed.

3) The learner is given a list of nodes from the computer’s map. Heading the list is the node currently calculated to be under discussion. Underneath is a list of nodes directly linked (on the computer’s concept map) to the first node.

Experimental results to date show that learners are able to integrate the textual enhancement information along with that obtained through the simultaneous voice conversation, creating a combined memory for the enhanced learning experience. The extra information allows the learner to be more proactive in directing the conversation, producing more effective learning conversation.

Selected references:
Rudman, P., Sharples, M., Baber, C. (2003) Automatic Support for Conversations using Semantic Relationships. Paper accepted for CAL 2003

Rudman, P., Sharples, M., Baber, C. (2002) Supporting Learning in using Personal Technologies. Proceedings of the European Workshop on Mobile and Contextual Learning. pp. 44-46. Available as PDF (127Kb).

See also my page at the University of Birmingham for a demonstration version of the software.

Note: This research has been made possible by joint funding from the Engineering & Physical Sciences Research Council and British Telecom

To Paul's home page"

(Last update 31-Jan-03)