SSNT - Summarization of Broadcast News Services

From HLT@INESC-ID

Revision as of 19:16, 12 February 2006 by Root (Talk | contribs) (System description)

This is the presentation of the SSNT service through a detailed description of his features. During this description you can discover several relevant aspects for a complete perception of this new service that we are offering. If you are not interested in such a detailed description and want just a quick view of the service we propose two different alternatives:

  1. See the first news of the last Telejornal in RealVideo format here
  2. Direct access to the SSNT service web page

If you are interested in a detailed description of the service we propose the following set of points:

  1. Goals
  2. Support
  3. Functional diagram
  4. System description
  5. User interface
  6. Present limitations
  7. Access to SSNT

Goals

Nowadays there is a significant need to deal with large amounts of multimedia information. With this service we want to develop a selective dissemination of multimedia contents, mainly of TV broadcast news. The use of advanced techniques for the processing of BN programs, through a segmentation and categorization process, made possible the access to the contents of the programs based on an individual definition of the user profiles.

Through this new service we made available the 8 o'clock news program of the main channel of RTP (Telejornal). The users are able to define which thematic areas they are interested and after the automatic processing of the program they receive an email with the news that fit to the requested domains. Be one of them and start using now this new service.

Support

This system was initially developed in consortium in the scope of the European project ALERT between INESC ID Lisboa, 4VDO and RTP. The developments of the large vocabulary continuous speech recognition system have been supported by the project POSI/33846/PLP/2000 financed by FCT.

Alert.jpg Inesc-id.jpg 4vdo.gif Rtp.gif

Functional diagram

In the next figure a functional diagram of the service is presented.

Functional-Diagram-ALERT.jpg

As we can observe from the functional diagram, the system analyses a generic multimedia document and based on the contents segment it in coherent blocks, through a video and/or audio segmentation.

When the document contains audio an automatic transcription is performed through a large vocabulary continuous speech recognition system. Based on the block segmentation and on the text inside each block, resulting from the transcription or because is only a text document, an automatic detection of topics is performed in each block, with the possibility of clustering together several blocks in homogeneous segments according to the topics contents.

With the multimedia document divided into segments, and a set of topics assigned to each segment, a search is performed on the user profiles requiring the topics from that segments and an alert message is generated for that users.

At the end of the process the multimedia document is loaded into a database where we keep the document segmentation and the appropriate categorization in topics.

System description

The development of the system was based on a three main blocks structure: the CAPTURE block, responsible for the capture of the monitoring defined programs, the PROCESSING block, responsible to generate the relevant markup information associated to each program, and the SERVICE block, responsible for the user interface and database management. The control of the overall process is based on a simple semaphore scheme.

Diagrama-ALERT.jpg

User interface

Registering as a new user

Reception of the system email

Direct search

Present limitations

This system presents a set of innovative features based on speech processing techniques and topic detection. However due to the development conditions we know that the system still have a set of limitations. Among them we highlight the following ones:

  • The speech recognition system is based on a limited vocabulary. Presently the system only have ability to recognize 58K different words. That means when there are new events establishing words out of vocabulary the system searches among the ones that are closer. This generates transcription errors with negative effects in the story segmentation and indexation.
  • Despite we are using different hierarchical levels in topics definition, not all the topics in a more deep level are perfectly trained due to the weak occurrence in the training process.
  • The title and summary do not have any processing involved. The title is only the first sentence of the news and the summary the first five sentences. This works reasonably good when the news is perfectly segmented and transcribed. In the future we want to make adequate processing to extract a title and a summary with a higher degree of correctness.
  • Since we are dealing with a new service is very important to have a very natural user interface. If you got any problems please send us an email with your suggestions.

Access to SSNT

After this detailed description you are in conditions to access to the system. We hope that you find in SSNT the necessary features to starting using this service. If you need any additional information or if you wish to draw any comment we are available here.

Access to the service
SSNT - Summarization of Broadcast News Services