IDENTIFYING NEWS EVENTS THAT CAUSE A SHIFT IN SENTIMENT

• US 20130290232A1
• Filed: 04/30/2012
• Published: 10/31/2013
• Est. Priority Date: 04/30/2012
• Status: Abandoned Application

• Current Assignee: Hewlett Packard Enterprise Development LP (Hewlett-Packard Enterprise Company)
• Original Assignee: Hewlett Packard Enterprise Development LP (Hewlett-Packard Enterprise Company)
• Inventors: Palpanas, Themis, Castellanos, Maria G., Dayal, Umeshwar, Hsu, Meichun, Tsytsarau, Mikalai
• Application Number:
  US13/460,541
• Publication Number:
  US 20130290232A1
• Time in Patent Office: Days
• US Class Current: 706/46
• CPC Class Codes: G06N 20/00 Machine learning
• CPC Class Codes: G06N 5/027 Frames

Abstract

A method identifies news events that cause shifts in sentiments. The method includes compiling a sentiment time series, the sentiment time series expressing a shift in sentiment; compiling a news events time series; correlating the sentiment and news events time series; identifying from the correlation news events that caused a shift in sentiment and predicting if a selected news event may cause a shift in sentiment in the future.

Claims

1. A method for identifying news events that cause shifts in sentiments, wherein the news events and sentiments relate to a same topic, the method, comprising:

compiling a sentiment feature time series expressing a shift in sentiment;

compiling a news feature time series expressing popularity/importance of news events;

extracting news event parameters from the news feature time series;

correlating the sentiment and news feature time series; and

identifying from the correlation a news event that caused a shift in sentiment; and

predicting if a selected news event will cause a future shift in sentiment

2. The method of claim 1, wherein compiling a sentiment feature time series comprises:

monitoring documents for sentiments;

detecting and collecting individual sentiments;

aligning the individual sentiments according to a time sequence;

aggregating the individual sentiments with respect to the topic;

determining sentiment feature values for the aggregated sentiments; and

identifying the sentiment shift in the sentiment feature time series based on the determined sentiment feature values.

3. The method of claim 2, wherein sentiment features include sentiment volume and sentiment contradiction.

4. The method of claim 1, wherein compiling a news feature time series comprises:

monitoring and detecting news stories as reported in news documents;

developing a news document time sequence for the topic; and

aggregating the news documents and extracting the news feature time series;

5. The method of claim 4, wherein extracting the news feature time series from the aggregated news documents comprises generating keywords'"'"' TF-IDF scores from the news documents.

6. The method of claim 1, wherein correlating the sentiment and news events time series comprises:

determining a time lag between the news feature time series and the sentiment feature time series; and

correlating the news feature time series with the sentiment feature time series.

7. The method of claim 1, wherein identifying a news event that caused the shift in sentiment comprises:

selecting the shift in sentiment;

navigating to a news event correlated in time to the shift in sentiment;

assigning news stories to the news event; and

creating a news event annotation.

8. The method of claim 1, wherein:

extracting the parameters comprises performing a deconvolution of the news feature time series to determine the news event parameters, wherein the parameters include time and longitude, buildup and decay rates and importance; and

predicting the shift in sentiment comprises using a classifier, wherein the classifier takes event parameters as input data, and uses event parameters with sentiment shifts as training data.

9. The method of claim 8, wherein:

the deconvolution is performed using one of a linear media response function;

mrf(t)=√

{square root over (2τ

₀)}−

τ

₀t, and an exponential media response function, mrf(t)=1/τ

₀·

exp(−

t/τ

₀), where τ

₀ is a time constant.

10. The method of claim 8, wherein:

prediction of the sentiment shift is performed using supervised machine learning methods, including Support Vector Machines, Decision Trees, Naï

eve Bayesian.

11. A system that identifies a news event that caused a shift in sentiment, the system comprising a processor having a program, the program, comprising:

a sentiment monitor that detects sentiments from multiple sources;

a sentiment aggregator that aggregates the detected sentiments;

a sentiment feature analyzer that generates a sentiment feature time series of the aggregated, detected sentiments, the sentiment feature time series expressing a shift in sentiment;

a news detector that detects documents that report a news event, wherein the news event is relevant to the detected sentiments;

a news feature analyzer that generates a news feature time series that expresses a measure of the popularity of the news event;

a time series correlator that correlates the sentiment feature time series and the news feature time series to identify if the news event caused the shift in sentiments;

a classifier model that predicts if the news event will cause a future shift in sentiments; and

an event describer, wherein if the correlation indicates the news event caused the shift in sentiments, the event describer annotates the news event.

12. The system of claim 11,

wherein the correlator determines a time lag between the time series according to one of a maximum of a cross-correlation coefficient:

max(|p(cf(t),nf(t−

τ

))|), and a list of most probable time lags, ranked according to a correlation coefficient; and

wherein the sentiments feature analyzer determines interestingness function values, comprising computing a sentiments contradiction value according to;

W(n)·

σ

_s/(μ

_s)².

13. A computer readable storage medium comprising program instructions that when executed by a processor, cause the processor to:

detect sentiments;

aggregate the detected sentiments;

generate a sentiment feature time series of the aggregated, detected sentiments, the sentiment feature time series expressing a shift in sentiment;

detect documents that report a news event, wherein the news event is relevant to the detected sentiments;

generate a news feature time series that expresses a measure of the popularity of the news event;

correlate the sentiment feature time series and the news feature time series to identify if the news event caused the shift in sentiments;

identify if the news event will cause a future shift in sentiments; and

annotate the news event, wherein the news event may be one of an event that caused or will cause a shift in sentiments, or an event selected by an operator.

14. The computer readable storage medium of claim 13, wherein the processor:

performs a deconvolution of the news event time series by estimating a time constant for news events sequence, and

estimates news event parameters.

15. The computer readable storage medium of claim 13, wherein the processor performs a prediction of the news event causing a shift in sentiment time series;

and wherein the processor, in identifying the news event, annotates the news event.