VAND Multi: Frontend Interaction - PMeschenmoser/Visual-Analyzer-for-News-Dependencies GitHub Wiki

Overall Interaction

  • Make sure to have both the app and Stanford Core NLP server running.

  • Open localhost:3000.

  • Now you can import articles (either .txt or .json) files from your filesystem. The text file name needs to correspond to a specific pattern (open the gear!).JSON files need to have a flat hierarchy and contain particular keys. These keys and the pattern for .txt can be modified within the settings.

  • It is also possible to import prototypical sets from the server. Simply search and select one or more table rows and apply it.

  • Now an initial 1:n comparison view is drawn. All imported articles are illustrated by the boxes on the left and horizontally ordered by their publication date.

  • Now we can compare the first article in time (c.f. full-text on the right, resp. the article in the blue box) against all others (-> reference articles).

  • A box is expanded if there is at least segment match between a reference article and the selected main article. Boxes for reference articles yield the segments in a compressed form, segments in the main article are highlighted in the full-text.

  • Given two (hovered) segments, the matching features are underlined.

  • Hovering an segment/edge: Highlight all belonging segment(-partner)s of this particular component.

  • Mouseenter on reference article: Highlight segment(-partner)s for all segments in this reference article. Plus: Display timestamp at the article and highlight corresponding entities in the context visualisation

  • Left-click on Article: Open 1:1 comparison view. Note that - as part of the overall interaction concept - dialogs may be deactivated within this view.

  • Right-click on Article: Set this article as new main article.

  • Right-click on Full-text segment: Bring this segment to the background. By these we provide a solution for overlapping features.

  • (Ctrl|Alt|Shift|Meta) + Right click on Article: Move this article to the back.

  • Zoom: By double-click, mousewheel or changing the size of the context bars.

  • Pan: By dragging the canvas or context bars

Settings

Heuristic Settings

  • MatchingFeatures[x,y]: All segments are extracted where at least x and at most y preprocessed and unique features are in between.
  • BetweenFeatures[x,y]: Segments may have between x and y non-matching, unique and preprocessed features in between.
  • OrderRatio[x]: All segments are extracted where feature ordering is kept for at least x*100 percent.
  • Matchtypes: All segments are extracted which have only features of the selected types, i.e. this is an AND filter. Currently, part-of-speech tags are possible match types.

Visualisation Settings

  • Border Width Encoding/tfidf cosine similarity: For each pair (mainarticle, reference article), we compute tf.idf based cosine similarities and encode similarity values to the border thickness for reference articles.
  • Border Width Encoding/Accumulated length ratio: Aka copy ratio. For segments in reference articles, we sum up the counterpart segment length in the main article and compare this sum to the main article length. By these mentioned metrics, we can already guide to suspicious reference articles.
  • Edge Selection Threshold: Highlight all edges, which lead to an reference article having at most a particular accumulated length ratio (cosine similarity).
  • Segment Opacity Factor: Purely visual parameter. It can be varied to be more flexible in terms of overlapping segments in full-text views.