Kanjoya’s natural language processing (NLP) technology and machine learning algorithms precisely and accurately understand human emotion, providing key insights into behavior and motivation

Emotion drives human decision making. Without emotion, it is simply impossible to understand and optimize the experience of your customers and employees. From customer and employee surveys to social media, Kanjoya analyzes any and all channels of data in real-time, delivering the insights that explain and improve business results.

Emotion Recognition and EPS

Kanjoya’s proprietary technology goes beyond simple positive and negative sentiment to decipher emotions, human factors--including promotion, detraction, and activation--and nuance in language.

Audience Categorization

We integrate emotion analytics with structured demographic information (e.g., location, age, and gender) to provide rich, high-signal guidance that improves business results.

Intelligent Baselines

We benchmark your data against the statistical profiles for different data sources, enabling you to identify ‘hot spots’, unusual activity, and key trends in your business.

Multi-dimensional Data

Built for massive scale, Kanjoya’s NLP technology automatically ingests vast sets of unstructured data, from Facebook and Twitter to surveys, search trends, and forum posts.

Kanjoya NLP technology is based on a decade of innovation and co-developed with leading linguistics researchers from Stanford University.

Moritz Sudhof, Chief Data Scientist at Kanjoya, manages the development of Kanjoya’s sentiment, emotion, and experience modeling technology.

His team, in collaboration with researchers at Stanford University, is leveraging the rich corpus of emotion data from Experience Project to model human expression in language.

Moritz discusses the world beyond sentiment analysis: context-aware emotion modeling for customer insights.

Perception Whitepapers

Kanjoya Perception vs. Sentiment Analysis

For Employee EngagementFor Customer Experience

Sentiment Analysis vs. Emotion Analytics for EE

Sentiment Analysis vs. Emotion Analytics for CX

Frequently Asked Questions
1. Is Perception different than sentiment analysis?

Sentiment Analysis relies on a simplistic classification system that only allows information to be designated as positive, negative, or neutral. These simple classifiers are often insufficient for anything but the most basic phrases. In contrast, Kanjoya Perception utilizes advanced natural language processing to identify a full range of emotions in unstructured text. Perception is able to handle nuance within polarity (positivity or negativity), conflicting polarity (the presence of both positivity and negativity), and promoter/detractor detection, and allows customers to understand and predict human behavior with a much higher degree of accuracy than sentiment analysis products.

2. What kinds of data does Perception analyze?

Perception can handle any type of unstructured (open-ended) text as well as data integrated from multiple structured and unstructured sources. Common data types for Perception customers include open-ended survey responses, reviews from sites like Yelp, Trip Advisor, and mobile app stores, social sites, such as Facebook and Twitter, enterprise networks like Yammer, and text from help-desk systems. For the list of Perception’s integrations, please see question 13. 

3. What is the setup time for a new Perception system? How much training is required?

Perception includes a consumer-based UI to encourage a broad range of users across the enterprise. Training requires less than an hour, and the setup of new data streams generally requires less then 2 minutes.

4. How much data can Perception process?

Perception is built to handle vast amounts of data and can scale to handle the largest enterprise deployments. Dozens of Fortune 500 companies already rely on Kanjoya to support their marketing, customer experience, and employee engagement requirements.

5. How much data is needed for meaningful results?

Perception customers obtain meaningful results from datasets ranging from 500 rows to as high as five million social media messages. In general, the larger the dataset, the better. 

6. Can Perception handle existing data or does it only work with new data?
 Perception handles both existing data–such as unstructured text from surveys or reviews–as well as new data, which can be collected through Perception’s complete survey solution or through other channels. 
7. How long does it take to calculate results?
Results are calculated almost instantaneously. In a matter of seconds or minutes, Perception delivers results for volumes of data that used to take weeks or months to analyze. 
8. Does Perception include spam filters?
Spam is context-dependent. What’s considered spam by one user (e.g., re-tweets of promotions) may be highly relevant to another customer. As a result, Perception handles spam by providing customers a way to filter across various dimensions, such as promoters and detractors in a customer experience or employee engagement context.
9. What is an EPS score? What is a good/bad score?

EPS stands for Emotional Promoter Score and is similar to a Net Promoter Score (NPS). Like NPS, the score identifies promoters and detractors in your data, and is calculated with the following formula: %Promoters -% Detractors. Unlike NPS, however, EPS does not require a special survey–Perception automatically generates an EPS for all of your data, and for any segmentation thereof! This provides constant monitoring of the promoters and detractors in your data, and an understanding of the topics and themes driving their behavior. 

Similar to NPS, good and bad scores typically vary across industries. In general, a good EPS is one that corresponds with an industry’s baseline NPS.  

10. Is there a relationship between emotions and promoters/detractors?
The models for emotion and promoters/detractors are distinct. Emotion models identify the emotions present in a statement, regardless of where/ how the emotions are directed. Promoter/detractor models, on the other hand, identify whether the speaker is promoting or detracting, from a particular company, product, culture, etc. The following statement helps to clarify the difference between these models: “I had a horrible day, but the people at the Ritz made it better.” In this case, Kanjoya’s emotion model recognizes that the speaker is upset, and the promoter/detractor model determines that despite the presence of negative emotions, this person is a promoter for the Ritz organization. Kanjoya’s promoter/detractor model is particularly effective in part because it handles emotions as an input, thus providing more meaningful results than most analytical systems. 
11. How does Perception handle both unstructured and structured data?
When Perception integrates structured and unstructured data, the structured data is retained as a set of filters that can be applied to the unstructured data. To make this more tangible, let’s consider the example of an employee pulse survey. As soon as a pulse survey is imported into the Perception interface, one is able to segment or filter open-ended text responses using any structured or demographic fields included in the survey, like gender, position, department, tenure, office location, etc.
12. What are the technology requirements to run Perception?
There are no installation or backend requirements in order to use Perception. All that is needed is an updated version of Firefox, Chrome, and Internet Explorer (IE 9 and above). 
13. Does Perception work with other platforms or products that are common in the enterprise, such as Medallia, enterprise social networks like Yammer, and in-house surveys?

Customers may import any CSV or Excel file, as well as Medallia and other internal data sources, including email.

Perception’s integrations for enterprise social include Facebook at Work and Yammer. Social media integrations exist for Glassdoor, Twitter, Facebook, WordPress, and Blogs. Perception also integrates with the survey products Survey Monkey, Qualtrics, and Perceptyx, and reviews from the Apple App Store and Google Play. Finally, Perception integrates with news and discussion sites including message boards, Reddit, Topix, and IntenseDebate, as well as IMDB.

Additionally, Perception can distribute its own surveys and collect responses through the complete survey solution.

14. How is Perception different from other analytics and sentiment analysis solutions?
Perception relies on an industry-leading natural language processing engine that is more accurate and better able to handle qualitative and emotional factors than its competitors. In addition, Perception handles not only existing structured and unstructured data; it also collects new data through the complete survey solution.
15. How were Kanjoya’s emotion models created?

Kanjoya’s emotion models are statistical machine learning models that are trained on millions of documents. Each document consists of unstructured text–a comment, an exclamation, a review, a statement–paired with an emotion label that describes the affective content of the text. Crucially, it is the author of the unstructured text that provides the emotion labels; as a result, they are as accurate as possible and are based on genuine social connections about topics, ideas, and feelings. These emotion labels are far superior to the post-hoc labels assigned to text through annotation services such as Mechanical Turk (people give minimal thought to labeling because they are not personally invested in the data).

In other words, our models are essentially trained on millions of instances of real people telling us “I said ‘X’ and I was feeling emotion ‘Y'”, and our concept of an “emotion” is empirical as opposed to theoretical or rule-based. Each training document we have is another observation of an emotion “in the wild”, and our concept of what “annoyed” means, for example, is the result of millions of people voting and telling us what they believe “annoyed” means.

16. How are Kanjoya's emotion models evaluated for accuracy?

As is common for machine learning models, we evaluate our accuracy through held-out test sets, which are documents that are representative of our application domain and are not seen by our models during training. Each document contains a label, provided by the author, which describes the emotion of the text. To compute accuracy, we compare our models’ predictions for emotions to the emotions labeled by the authors of our test sets (please see FAQ #15 for more information regarding the creation of these models).