Sentiment Analysis

Sentiment Analysis assesses the opinion expressed in content, creating a quantified measurement of the author's point of view. Attivio's Sentiment Analysis can be used to evaluate many types of content, such as web postings, emails, comments added to surveys, content in call logs. Once analyzed, the sentiment data can be used to report individual satisfaction or assess a group of opinions, such as overall market view of a product. This sentiment information can be reported as numeric values or terms such as "mostly positive" that are based on numeric ranges that you define.

Attivio Active Intelligence Engine® (AIE®) includes both document-level and entity-level sentiment analysis. Document-level sentiment analysis determines the overall opinion expressed in the document, while entity-level sentiment captures views about an entity that is mentioned in the document. The entity-level sentiment may be quite different from the overall sentiment. For example, a group of product reviews may be generally positive, but also contain negative comments about one particular feature. Entity-level sentiment is useful for staff such as product managers who need to track the specific aspects of a product or service that is either a strength or weakness.

Why Sentiment Analysis Is Important

Because the internet provides so many opportunities for people to publish their views, companies are under increasing pressure to track and analyze opinions and attitudes about their company and its products and services. These opinions from customers, prospects, reviewers, analysts and employees are expressed in various web forums, from consumer review sites to Twitter as well as in emails, call logs, and web-based surveys and forms. By monitoring and analyzing opinions, companies can gather intelligence on a customer's defection risk, the impact of an influential reviewer on other people's purchase decisions, and level of satisfaction or intensity of complaints about products and services and about the company and its competitors. Capturing and analyzing these opinions is a necessity for proactive product planning, marketing and customer service. It is also critical in maintaining brand integrity.

The challenge for leveraging sentiment is tracking disparate sources and then accurately capturing the meaning in the opinion in time to effectively analyze and act. Opinions are expressed in many different ways; accurately analyzing and measuring this diverse content produces quantitative values that improve the usefulness of the data companies rely on to run their business. In addition, entity-level sentiment helps refine the data to give insight into specific topics within documents, allowing deeper and more accurate analysis.

Basics of Sentiment Analysis

Sentiment analysis extracts and measures the sentiment or "attitude" of documents as well as the topics within documents. The attitude may be the person's judgment (e.g., positive vs. negative) or emotional tone (e.g., objective vs. subjective). Sentiment analysis can enable companies to:

• Quantify individual and aggregate opinions (e.g., 27% of consumer comments about our new product are negative; 90% of complaints are about battery life)
• Gauge the satisfaction level of a single customer or a group of customers (e.g., using a coded ranking of which statistics map to defined satisfaction categories, AIE can report that recent call log and email content show a customer falls into a "somewhat dissatisfied" category)
• Evaluate sentiment about multiple topics in a single document (e.g., "The new model is easy to use and well designed, but performance is too slow.")

Once this opinion is quantified, you can more easily use it in decision making and customer service. For example, sentiment data can be automatically forwarded to or queried from applications such as a dashboard for Voice of the Customer, emailed to designated people who need to be alerted, or added to a data store for use in project planning, product design, performance management, etc.

How Sentiment Analysis Works

Language is complicated and contextual, especially when people are expressing opinions. How does AIE accurately analyze something as non-quantitative as sentiment, which can be expressed in many different ways, using terms whose meaning can vary ("nightclub XYZ is hot" does not mean the same thing as "the battery gets too hot")? For sentiment analysis to be effective, it must also understand context and topic-specific terminology.

In sentiment analysis, AIE classifies the opinion expressed in content as positive or negative by applying an algorithm that detects and evaluates the point of view expressed. That assessment is rendered as a quantitative value that is indexed along with the content. This means AIE adds structured information (the quantitative assessment) to unstructured text containing an opinion found in documents, blogs, call logs, emails, etc.

To create the structured data, AIE does two things:

1. Performs analysis of opinion, giving the opinion a quantitative value and a degree of accuracy
2. Adds this metadata information to the index

One benefit of Attivio's sentiment analysis is that in combination with entity extraction*, users can investigate details about the opinions to determine not just the overall point of view but also the nature of both positive and negative comments. For example, a product manager researching why 32% of product comments and reviews are negative can query the extracted entities in the reviews to find out that most of the negative comments were about a single feature, poor packaging, etc.

* AIE extracts and indexes entities (names, noun phrases) in all ingested content, regardless of whether sentiment analysis is also applied.

Training the System

Different topics tend to be discussed with distinct patterns and types of terms, so one important aspect of all sentiment analysis tools is one-time "training" of the system to establish the typical terms used in that context. For example, when you are implementing Sentiment Analysis for a market intelligence application used by an electronics firm, you would train the system to understand that context. AIE applies a sentiment model for each category when it analyzes sentiment.

All that is required for training is to identify some documents that are good examples of positive and negative sentiment in the context in which sentiment analysis will be applied. The AIE model builder creates a sentiment analysis model from these documents. As the model is built, the model builder also reports accuracy based on cross-validation, a standard machine-learning technique. To achieve the highest accuracy, you should ingest at least 1000 documents, which takes less than a minute.

After training on a sample set of documents, AIE has learned what types of sentiment language are used in a related group of documents (e.g., groups of performance evaluations, groups of customer calls, groups of product reviews, etc.). After the initial training, you may also need to conduct some analysis and some fine tuning to ensure highest accuracy.

Applying Analyzed Sentiment

Sentiment analysis is fully integrated with the rest of AIE. After you train the system and begin collecting sentiment data, you can query for any aspect of the data, graph results over time or examine comparative sentiment. As always with AIE, you can use saved queries and alerts to ensure timely sentiment monitoring and notifications, including alerts when opinion volume has changed beyond a certain threshold or when an especially positive or negative sentiment has been published. You can also retrieve sentiment data for display in dashboards or for use in formulas for assessments.

Besides actively monitoring consumer and thought leader opinions online, there are other significant uses for sentiment analysis. Companies can use sentiment analysis to enhance the value and actionability of employee surveys, IT trouble tickets, and performance evaluations. Sentiment analysis is one more way unified information access delivers timely and complete information on demand so people can gain deeper insight and make better decisions.