Hacking Emotions: Interpreting Current Events through News Sentiment

News sentiment – the opinions, emotions, or attitudes presented in the news – can play a significant role in customer and investor behavior. As a result, brands and organizations may want to know the sentiment regarding a particular product or a relevant event.

They may also want to discover any correlations between an article’s virality and the different entity sentiments in it. Financial institutions can also use news sentiment to gain insight into future market movements.  

How can organizations best measure news sentiment to gain these types of insights? 

News sentiment analysis is typically measured on two levels:

* Document-level sentiment - the sentiment of a news article as a whole

* Entity-level sentiment - the sentiment associated with an entity (i.e. a person, place, or thing) within a news article

Sometimes a holistic view of the news sentiment of an article isn’t sufficient. An article might be positive on the whole, while most of the entities mentioned could have an associated negative sentiment.

The article-level sentiment can also affect the entity-level sentiment. Different entities can also have widely different sentiments. For example, an organization’s brand name might have a positive sentiment while its stock ticker might have a negative sentiment.

Let’s take an example search query using our Enriched News API for articles mentioning Tesla’s Cybertruck with an overall positive sentiment polarity. Within these articles, other associated entities may have a neutral or negative measurement.

The article is shown on the left below, about a tweet announcing the new Tesla Cybertruck has a positive sentiment for the article as a whole, but it detects a negative sentiment for “Elon Musk”; a neutral sentiment for his Twitter handle “@elonmusk”; a neutral sentiment for “lorry” (truck in British English) and positive sentiment for “Twitter”!

Other articles about the Tesla Cybertruck have a different set of sentiments for the same entities. In the result on the above right, “lorry” is neutral, and “Elon Musk” is positive. 

Here is another query for articles with a positive sentiment polarity about the stock ticker TSLA.

As the examples below show, the article on the left about Tesla’s stock price has an overall positive sentiment, but a negative sentiment for “stock.” The article on the right, however, has an overall positive sentiment, but a positive sentiment for “stock.”

Again, other articles on the same subject can detect different sentiments for the same entity. The article on the above right about Tesla’s stock has an overall positive sentiment for the article and detects a positive sentiment for “stock.”

Another important point to keep in mind is that an article was written with irony and/or subjectivity can influence the entity sentiment. For example, if Elon Musk receives a negative sentiment, but the article is written in an ironic or subjective tone, it’s not the same as if Elon Musk received a negative sentiment in a more objective news article.

The article above has an overall negative sentiment, but a positive sentiment for “Elon Musk.”  It was written for a publication about pop culture and does not have an ironic tone.

All of these different entity-level search results are interesting, but analyzing articles individually doesn’t demonstrate the full value of entity-level sentiment. It’s when thousands of articles and entities are aggregated at a scale that the real magic happens. 

Here is a partial list of entities detected from a search for the topic of “Olympics” and their corresponding positive measurement in 9,547 articles.

Interestingly, athletes qualifying for the United States Olympic basketball and wrestling teams had the highest entity sentiment scores averaging around 80%. The scores were slightly higher for male wrestlers than for female wrestlers, but not by much. 

If we compared a list of the number of times “male” and “female” appeared in the different articles and their corresponding sentiment, would there be a significant difference? Can we attribute this difference to a preference for male athletes over female ones? 

Another highlight we found was that soccer athletes qualifying for the US team had lower positive sentiment scores ranging from 58% to 70%. Can we correlate this result with how Americans perceive soccer as a sport? In general, do more experienced athletes, who would also get more mentions in the media, have a tendency to have a higher positive sentiment? 

Finally, if we drill down according to the author or reporter of different articles, would we see that articles by the same reporter on a certain sport or about particular athletes have a higher or lower sentiment? Does this sentiment reflect an accurate public opinion of those sports or athletes? 

These are all types of insights that entity-level sentiment can help deliver.

News sentiment analysis at the entity level delivers a more granular viewpoint. This is particularly valuable when looking at thousands of articles and entities to determine patterns and trends. Financial institutions can then apply these insights to their predictive models. Brands can start to distinguish between changes in brand or product sentiment. Organizations conducting risk intelligence can more easily detect adverse media. 

Webhose’s Enriched API can give these organizations the high-quality news sentiment analysis they need – on both the article and entity level.

Previously published here.