About sentiment analysis

Talkwalker's AI-powered sentiment analysis provides a way to analyze the emotional tone of a piece of content, such as a social media post or an article. You can use the information you gain from sentiment analysis to streamline workflows, tailor responses based on emotional urgency, and help inform strategic decisions and crisis management strategies.

About the sentiment analysis model

Talkwalker’s sentiment analysis currently recognizes 186 languages and can provide an overall accuracy of up to 90%. Talkwalker's AI-powered sentiment analysis model combines the following advanced techniques to get the best results:

• Deep learning - We do not train our engine using our customer data. It uses a predefined dataset of tens of millions of human-annotated results to help it identify relevant patterns and improve results.
• Relevant patterns - Our algorithms, which run behind the scenes, don't just decipher individual words, they can understand the meaning behind complete sentences. This allows them to interpret human emotions, basic types of irony, and sarcasm.
• Neural networks - We employ powerful computing methods that simulate human cognitive skills.
• Natural language processing (NLP) and machine learning - We use these techniques to assist in determining sentiment, to recognize image and video, and to assist in machine translation, Theme cloud creation, demographics, and Conversation Clusters.

Note: AI sentiment models are not trained on customer data.

How we assign sentiment

For features that support sentiment analysis, you can choose to use AI to automatically assign sentiment and you can manually change sentiment. AI analyzes whole sentences (not just key words) and includes all results to determine the overall contextual attitude or reaction. Only one sentiment can be assigned to each article or post.

Sentiment calculations in Inbox 2.0

We use TalkWalker’s AI service to determine positive, negative, or neutral sentiment for individual messages in a conversation. We also calculate an overall sentiment for each conversation. We use the last five non-neutral messages in a conversation to calculate the sentiment for that conversation. For each positive message, we add one point and for each negative message, we subtract one point. We then average the last five scores and use the following range to calculate the conversation sentiment:

• Positive - Between 0 and 1.
• Neutral - 0.
• Negative - Between -1 and 0.

For example, we calculate a conversation with 1 negative, 1 neutral, and 3 positive messages as positive. (-1 + 1 + 1 + 1 )/4 = 0.5, which is between 0 and 1 and is therefore positive.

Sentiment calculations in Talkwalker

If AI Classifiers are a part of your package, you can train custom sentiment AI Classifiers for your project.

AI sentiment is applied to results manually imported into Talkwalker. However, if you indicate your own sentiment in the import file, the automatic sentiment will not be assigned.

Global AI sentiment is content-based, not brand-based. For brand-based sentiment, use custom AI Classifiers.

Emotion detection

Both sentiment and emotion are detected via AI. However, they are calculated separately and different methods are used for their calculations.

For emotion detection, we follow Parrott’s structure of emotions. Emotion detection is based on individual sentences and one article can appear in multiple emotion categories. Emotion detection only applies to results in which emotions could be identified.

The emotions detected for a specific result may not necessarily match the sentiment of the result. For example, joyful emotions in a post do not necessarily mean the post will be classified as positive.