rianne nieland's final presentation

18
Talking to Linked Data: Comparing voice interfaces for general- purpose data Rianne Nieland Supervisor: Victor de Boer Vrije Universiteit Amsterdam

Upload: victor-de-boer

Post on 19-Aug-2014

178 views

Category:

Education


8 download

DESCRIPTION

These slides are made by Rianne Nieland and were presented by her to finalize her Master project Information Sciences

TRANSCRIPT

  • Talking to Linked Data: Comparing voice interfaces for general-purpose data Rianne Nieland Supervisor: Victor de Boer Vrije Universiteit Amsterdam
  • Context & Problem Statement Web = big information space o Contains useful information for people in developing countries Like governmental and medical information, and information about plants and trees on Wikipedia People in developing countries: o No internet access o Often low literate o Do have mobile phones Solution: Voice-based access to Web data using GSM network Research: Develop voice interfaces for general-purpose datasets
  • Wikipedia vs DBpedia Natural language text And structured information, like infobox, images and links to other pages Extracts structured information of Wikipedia DBpedia ontology: classes and properties Data interlinked with other data sources Very lightweight way to share, re-use and integrate datasets
  • Research Questions How can information from Wikipedia efficiently be made available using voice interfaces for GSM? 1. What are the requirements of a good voice interface for Wikipedia and DBpedia concepts? 2. What are good methods for converting Wikipedia and DBpedia concepts to voice interfaces? 3. How do users perform on the Wikipedia and DBpedia voice interface in terms of speed, error rate and usability?
  • Approach Experiment Developing conversion algorithms (Process) Developing voice user interface (Output) Requirements elicitation Literature study
  • Requirements elicitation Input requirements o Dual-Tone Multi-Frequency input o Local phone line Process requirements o Overview of page o Eliminate repetitions o Feedback o Error recovery Output requirements o Systems voice: female + text to speech o Nonverbal sounds
  • Voice user interface Basic call flow structure: 1. Welcome message + page menu 2. Section menu 3. Subsection menu 4. Reads chosen (sub)section to user Voice interfaces have same basic call flow structure But different input sources
  • Process of voice interfaces Input: Wikipedia /DBpedia ? Output: Call flow
  • Process of voice interfaces Input: Wikipedia /DBpedia Proces: Conversion Output: Call flow
  • Conversion steps DOMXPATH queries Section menu: o Elements with class mw- headline, except h3 and h4 Subsection menu: o h3 elements Read (sub)section o p en li elements SPARQL queries Section menu: o Abstract o Nutritional values o Biological classification o Associated food, persons and organizations Subsection menu & Read (sub)section o SPARQL queries
  • Experiment 16 participants Domain crops Each participant tests both voice interfaces by answering questions with the voice interfaces 2 question sets of each 3 questions Divided participants into 4 groups: o First Wikipedia (W) with question set 1 and then DBpedia (D) with question set 2 (W1D2) o W2D1 o D1W2 o D2W1
  • Experiment 1. Verbal explanation 2. General questionnaire (gender, age, purposes of mobile phone usage and usage of voice interfaces) 3. Test first voice interface by answering questions 4. Fill in IBMs usability satisfaction questionnaire 5. Test second voice interface by answering questions 6. Fill in IBMs usability satisfaction questionnaire
  • Results: Speed Wikipedia voice interface average time 2:53 minutes DBpedia voice interface average time 2:22 minutes No significant difference Both voice interfaces equally fast Also no learning curve found
  • Results: Error rate In general both voice interface have significantly the same error rate For question 2 of question set 1 Wikipedia has a significantly lower error rate
  • Results: Usability Usability is divided into 4 scores: o Overall satisfaction o System usefulness o Information quality o Interface quality In general no significant difference between Wikipedia and DBpedia voice interface for all scores When voice interfaces are tested first: o Wikipedia scores higher on overall satisfaction and information quality
  • Discussion WiFi connection sometimes did not work DBpedia was offline a number of times o DBpedia backup voice interface Participants are used to access textual version of Web DBpedia contains very little information DBpedia voice interface is domain specific
  • Conclusion To make information from Wikipedia efficiently available using voice interfaces for GSM: o Requirements should be met o Conversion methods used in this research should be considered, because they work efficiently o Both normal Web data, Wikipedia, or Linked Data, Dbpedia, can be used
  • Future work Can be used in developing countries o Should use local languages o Local phone number o Should be tested there outside a lab o Investigate what information these people need Broaden scope to whole Wikipedia and DBpedia