PASCAL has been very successful in supporting and promoting the development of the videolectures site (www.videolectures.net). Knowledge 4 All Foundation has been established to carry forward the development of this activity as well as other legacies of the PASCAL Network. One problem created by the success of videolectures is the difficulty that individual users have in identifying the best video for their needs among the vast range of possibilities afforded by the site. Each video has a particular mix of content and style of presentation with implicit assumptions about background knowledge and level of expertise of its intended audience.
On the other hand the video consumer has an approximate understanding of his/her abilities and material that he/she would like to learn about. For example, they may have a background in basic classiffication methods (eg SVMs) applied to text, some knowledge of probability theory, but not know about Bayesian reasoning. He/she would like to learn about Topic Models. The question of which sequence of videos would be most appropriate to help him/her to attain the desired knowledge would also depend on the style of presentation he/she prefers and so on. Currently Videolectures provides a contextualisation service that is only based on keywords extracted from the lecture titles.
It provides a recommendation of videos that are related to a given video. The relation would typically be based on the topic of the video or the lecturer. Furthermore, the system does not adapt its responses to the interests or background of the user. Videolectures has begun to collect information about individual users, though it is currently limited to associations with the keywords of lectures that they have viewed. This information is currently only being used for o-line analytics.
Proposed Harvest Project
The La Vie project would develop a proof-of-concept system that would provide users with advice on suitable videos for their needs. The key additional components that the project will bring to videolectures are:
- topic extraction and modeling based on text extracted from associated slides and audio transcriptions. This will ensure that the devised user models can capture semantic level interests of the users.
- inclusion of the information currently being logged about individual users in the recommender system running live on the videolectures site.
- visualisation of the developed recommender system. This will provide a topic landscape that will enable users to see the available videos emphasising those likely to be of interest to the user.
The La Vie system will be broken down into ve phases. First the text extraction modules will use text mining methods to extract information from the content and meta-data associated with a particular video. The second phase will involve topic modeling from the extracted text in order to develop a richer semantic representation. Phase three will use the developed topic models and other relevant information (such as formalised ontologies, taxonomies, structured information sources such as wikipedia, linked open data (LOD), and other contextual information) to populate a semantic representation of individual videos and users. The fourth phaseinformation retrieval module will provide relevant recommendations by linking the enriched semantic representations of users and videos. The nal phase collects feedback from users about the operation and e ectiveness of the recommendations that have been given. This information will be used to update the user and other models. All of these components will be integrated into videolectures through the user interface that will provide visualisation and interactivity.
More detailed descriptions of these components are given below. The project will provide a framework into which new modules can be plugged either to replace old ones or to enhance functionality. For example, a simple k-means algorithm would be used initially for the topic modeling, while clearly more advanced methods should provide improved performance and may be included in the system during the rst phase of the project if time permits. As an example of enhancing functionality the representation of user interests as a distribution over meta-data attributes would make it straightforward to include new modules re ning these distributions through interactions with a user applying say a bandit algorithm inspired by content-based image retrieval.
This framework approach to the development of the La Vie project ensures that the risks associated with the project are kept to a minimum. Once the framework is in place and the basic system operational, further incremental developments can be included without risking the overall success of the project results. Several components already exist including Qminer for user modeling, Enrycher for text enrichment, DocumentAtlas for visualisation, etc., that will make this approach likely to deliver a basic working system within approximately six weeks given the manpower envisaged below. Here we have taken into account the time required for the team to accumulate the necessary expertise in the di erent existing modules as well as the signi cant effort required to design and program the new framework and modules. In the second part of the project speci c advanced modules will be developed chosen based on the skills of the participants and potential for enhancing the overall performance of the system. These modules could include probabilistic topic models, bandit style algorithms for optimising user interaction, improved visualisation algorithms, etc.