Describing Time Series in Words

1
Describing Time Series in Words
V. Shiv Naga Prasad

• Generating English Summaries of Time Series Data
  Using the Gricean Maxims - Somayajulu G. Sripada,
  Ehud Reiter, Jim Hunter and Jin Yu

2
What not to do

• MANAGEMENT SPEAK The procedure which you have
  referenced is the correctly referenced procedure
  for the procedure which you referenced and which
  I am referencing in my reply about the procedure
  in which your original inquiry was referenced.
• TRANSLATION Yes.
• -- IS Survivalist Bill Murray swears he received
  this response from a not-to-be-named standards
  body

3
Generating Textual Summaries

• What is to be told?
• select salient trends in data.
• How to express it?
• Lingua franca of the domain.
• Generate the text.

4
Example of Issues

• Many things can be described.
• Many ways of describing.

5
What is significant?
Difficult question
This is actually an artifact and should be
suppressed during summary.
6
Gricean Maxims of Communication

• Quality stick to facts.
• Quantity neither too less nor too much.
• Relevance for the user.
• Manner be concise and consistent.

7
Approach

• Segment sequence into linear segments.
• Describe end-points of segments.

8
Quality

• Actual values better than estimates (quality).
• Use range to show uncertainity.

Line of best fit
Estimated value
Actual value
01-06 0.5-1.5 Increasing 4.5-5.5
9
Quantity
Eliminate artifacts maybe with interpolation
10
Relevance

• Small changes might not be important.
• This influences
• Segmentation.
• Scale of description.

11
Manner

• Be consistent use standard adjectives and
  ordering.

01-06 0.5-1.5 Increasing 4.5-5.5 06-09 4.5-5.5
Steady 09-13 4.5-5.5 Decreasing 0-0.5

• Avoid obscurity omit eventless segments.

01-06 0.5-1.5 Increasing 4.5-5.5 09-13 4.5-5.5
Decreasing 0-0.5
12
Evaluation Results for SumTime-Mousam

• Facts were correctly described.
• Humans eliminated much more irrelevant facts.
• System still rookie in terms of linguistic
  elegance
• Connectives (e.g. thenand).
• Repeated use of and.

13
Comments

• Currently describes linear trends
• how about specific patterns?
• Differentiating between anomalies and artifacts
  is difficult.
• Is textual description better than graphics?
• Using text-to-speech and textual summaries
  audio can be used as an additional modality.
• For visually impaired, summaries in brail would
  be useful.