Discussion Leader:
Tianyi Li
Summary:
Youtube Video:
An important part of visual analytics is the statistical models used to transform data and generate visualization (visual metaphor). Analysts manipulate the visualization by tuning the parameters in those statistical models in their information foraging process. When it comes to information synthesis, analysts most often externalize their thought process by spatial layout the information. However, there is a gap between the two processes in terms of the tools available, with most visual analytics tools only focus on one side or another.
This paper takes this opportunity and bridge the two processes via “semantic interaction”, which manipulate parameters of the statistical models by interpreting analyst’s interaction (mostly spatial) with the visualization. Such interaction relieves analysts from understanding the complicated statistical models so as to focus more on the semantical meanings of the text data under analysis. This also opens new design space for visual analytic interaction. The author demonstrated an example visual analytic tool they developed as a proof of concept.
The visual analytic tool, called ForceSPIRE, combines machine learning techniques and interaction analysis, to assist analysts in analyzing intelligence analysis text documents. The system focus on the document similarity, represented by entity overlap and the amount of user interaction happened (conceptually) and spatial layout (visually). User interactions will influence the values of encoded feature labels of documents, thus trigger updates on the visual layout to reflect their sensemaking progress.
ForceSPIRE encodes text documents via the following concepts:
- Soft data: stored result of user interaction as interpreted by the system
- Hard data: extracted from the raw textual information (e.g., term or entity frequency, titles, document length, etc.).
- Each entity has an importance value
- initialized with tf-idf score
- updated by user interaction hit
- always normalized to 1
- Each document has a mass
- The number of entities corresponds to the mass of each document a.k.a nodes in the force-directed graph model, where heavier nodes do not move as fast as lighter nodes.)
Reflection:
This paper is very well written and one of the cornerstone papers for my research area and project. Semantic interaction bridge the gap between information foraging and synthesizing, which relieved the burden of analysts to learn about complicated statistical models. In other words, it realized the purpose of those models, and a fundamental mission of visual analytics: to hide the unnecessary complexity of data processing details from the analysts so that they can achieve productive analysis via visual metaphors.
The way ForceSPIRE computes similarity is to start with “hard data”, which gives an initial spatial layout of documents. Then while users interact with the visualization, the result of user interaction (“soft data”) is stored and interpreted by the system to update the spatial layout in real-time. Entity is the smallest unit of operation and computation. The initial entity values are assigned by computing tf-idf scores. Then one hit of user interaction, i.e., the entity was included in a highlight, it was searched, it was in a note, etc. , will increase 10% of the value (thus reducing other entities importance value). Each entity has its importance value and the sum is always normalized to be 1. Document, which contains entities, is of one level higher granularity. They assign a mass to documents to account for the number of entities in it. Similarly to entities, documents also increase their mass by 10% with 1 hit of user interaction, i.e., text was highlighted, it was the result of a search hit, or if user added more entities through annotations.
This is an over simplified scenario just to prove the concept, and there are following works on this, like StarSPIRE, which released the assumptions of large screen and small number of documents.
Questions:
- In ForceSPIRE, it supports “undo”, where user press ctrl-z and rewind to the previous state without the most recent interaction. The entity importance values and document mass values are all rewinded but the spatial layout is not. They also recommended the users to have small-distance document movement. Why not enable undo for document location as well? What is the reason for not doing this?
- In ForceSPIRE, they focused on document “similarity”. In the second paper today, they analyze news content by “relevance” and “uniqueness”, which seems to me is a finer tuned breakdown of the concept of similarity: relevance is the similarity to the topic of interest, uniqueness is the negate of similarity. Do you agree with me? How do you think other “computational enablers”, borrowing the term from the second paper, can be used in semantic analysis?
- In ForceSPIRE, semantic analysis is applied to intelligence analyses which has specific facts to discover or questions to answer. How do you think semantic analysis can be applied to journalism or news content analysis? How are these scenarios different? How would such difference influence our analysis strategy?
