This article will cover:

Sentiment is defined as a view of or attitude toward a situation or event, an opinion.

Sentiment analysis analyzes a body of text to determine its opinion. The opinion is then labeled either more Positive or more Negative. This is called Polarity.

The overall sentiment is then determined to be either Positive, Negative, or Neutral based on the total polarity score. If something is both positive and negative, but more positive, the polarity score would rank it as positive.

Our sentiment model leverages transformer technology, a world-class, deep-learning

architecture. Transformer-based sentiment models enjoy some of the highest accuracy scores in the field of natural language processing.

Our latest model incorporates features specifically to ensure high-level accuracy:

The attention mechanism - deriving understanding by focusing on the most important and relevant words in the text.

Sentence-based embeddings - Our new model is adapted from a pre-trained, open-source model built by Google (BERT). Whereas old models represent words as vectors, BERT represents entire sentences as vectors. This means that our model is better equipped to handle words with multiple meanings. For example, mouse. Here are other examples:

Sentiment shows the overall tonality of any given article or mention. This provides a look at each result and consolidates the message's tone into either positive, negative, neutral, or not rated. These are derived from our natural language processing (NLP) algorithms. It helps give context to the result and, holistically, provides an overview as to the tone of your search, your brand, or relevant topics to you.

We support language detection for 242 languages and provide full sentiment Analysis for 218 languages and dialects.

Below are a few examples of Sentiment applications within Meltwater.

Positive. Here, a user is exclaiming love for a Coca-Cola campaign.

Negative. This user is displaying negative emotion towards the taste of a Coca-Cola product.

Neutral. This user is simply stating a fact about a Coca-Cola collectable, with no emotion implied.

Now for a tricky one. NLP and Sentiment Analysis is highly accurate, but not perfect. Slang, sarcasm, inference, and subtext, can all alter the way you perceive sentiment versus someone else. Two people might even look at the same message, and have differing opinions of sentiment! Here is a good example:

This user mentioned they like Pepsi, which would be positive for Pepsi.

They also mentioned they like Coca-Cola better than Pepsi. That is positive for Coca-Cola and negative for Pepsi. But they said they liked Pepsi too, so is that still positive for Pepsi?

This user mentioned Pepsi and Coca-Cola, but not RC Cola. Would this be negative, positive, or neutral for RC Cola?

We use this example to highlight that even with NLP, context still matters when determining sentiment as it relates to the overall goals within your results. Sometimes you may not agree with an assigned sentiment, and that is okay! Read the next section to see how to Override Sentiment within Meltwater.

With Sentiment Override, you have the ability to assign or change a sentiment assessment of a document or mention.

*Important Notes:

You can override sentiment in the content stream with bulk selection instead of overriding each result individually.

Note: If you select more than 500 documents or select the Select All box at the top of the content stream and click Select All Total Results the sentiment emoticon will disappear as no more than 500 results can be changed at once.

Within Classic Dashboards there are two widgets that can help you understand the overall tonality of a group of articles, these are named 'Sentiment' and 'Sentiment Score'.

Sentiment - This provides a look into the volume of documents assigned a tonality score of positive, negative, or neutral. It helps give context to the article and, holistically, provides an overview as to the tone of your search, your brand, or relevant topics to you. This graph can be displayed in the following chart types: donut (as seen below), line graph, solid bar, or pie.

Sentiment Score - It sums the percent of positive articles against the negative articles, ignoring 'neutral' articles.

Sentiment Compared to last period - This metric provides an overview of the sentiment of the results. It is computed by deducting the results with negative sentiment from the results with positive sentiment, then comparing this with the total documents with a sentiment rating. Formula: ([number of positive documents] - [number of negative documents]) / ([number of documents with a sentiment rating]) x 100

Pro Tip If half of your articles are positive, and the other half are negative, expect a score of '0'. Whereas if all are positive, expect a score of '+100'.

The following Sentiment widgets can be found in the Canvas Layout, Slide Layout, Brand Report Template, Benchmark Report Template, Campain Report Template, Coverage Report Template, and Earn Media Measurement dashboards.

Average Sentiment - Shows the average tonality of mentions across your selected inputs. Sentiment is derived from our natural language processing algorithm.

Sentiment Breakdown - Shows a sentiment breakdown for mentions across your selected inputs. Sentiment is derived from our natural language processing algorithm. Mentions labeled as "Not Rated" do not have enough text to analyze sentiment. YouTube data is excluded from this visualization per their policies.

Sentiment Trend - Shows a sentiment breakdown for mentions across your selected inputs, trended over time. "Not Rated" mentions don't have enough text to analyze sentiment. YouTube data is excluded from this visualization per their policies.

Sentiment by Source Type - Shows a sentiment breakdown for mentions across your selected inputs, separated by source type. "Not Rated" mentions don't have enough text to analyze sentiment. YouTube data is excluded from this visualization per their policies.

Average Sentiment Trend - Shows an average tonality of mentions for each of your selected inputs over time. YouTube data is excluded from this visualization per their policies.

Top Keywords with Sentiment - Shows the keywords and phrases that appear most often in mentions across your selected inputs, broken down by the sentiment of the mention. Mentions labeled as "Not Rated" do not have enough text to analyze sentiment. YouTube data is excluded from this visualization per their policies.

Top Publications with Sentiment - Shows the editorial publications with the most mentions, broken down by the sentiment of mentions. Mentions labeled as "Not Rated" do not have enough text to analyze sentiment. YouTube data is excluded from this visualization per their policies.

Top Locations with Sentiment - Distribution of mentions by location based on the location of the author, additionally split by sentiment. YouTube data is excluded from this visualization per their policies.

Average Sentiment - Shows the average tonality of mentions across your selected inputs. Sentiment is derived from our natural language processing algorithm.

You can now search by entity-level sentiment!

Using the new boolean for granular sentiment, you will be able to retrieve articles and posts where a certain entity (e.g. your brand, competitor, or product) is mentioned in a negative or positive sentence, while the overall sentiment of the document may be different.

For example, this query will return all documents where Tesla was identified as an entity and appeared in a negative sentence:

enrichments.namedEntities[canonicalName:Tesla AND sentiment.discrete:v]

To find positive results, change the sentiment variable from v (negative) to p (positive), like this:

enrichments.namedEntities[canonicalname:Tesla AND sentiment.discrete:p]

To cover multiple versions of the entity, you can expand the name section to an OR query:

enrichments.namedEntities[canonicalName:Tesla OR name:“Tesla Motors“ AND sentiment.discrete:p]

Sentiment boolean variables are

*Important Notes:

In instances where two or more entities are mentioned in the same sentence with different sentiments, the sentence sentiment will be assigned based on the majority combined sentiment.

For example: Coca Cola is the best, I don‘t like Pepsi.

This sentence is more positive than negative, so all detected entities in this sentence will receive a positive sentiment.

Entity searching is not yet available for social channels where a brand is mentioned as @handle, rather than with a name written out, such as Twitter.

We continue to add more languages to our NLP and modern deep learning models regularly so that we can continue to improve accurate sentiment detection and retrain languages when needed.

💡 Tip

Need more help? Feel free to reach out to us via Live Chat or check out our Customer Community.

Find answers and get help from Meltwater Support and Community Experts.