How to use modern language models for improved analysis of feelings

The analysis of feelings promised to reveal the secrets of customer comments, but obsolete methods have left us superficial information and too simplistic graphics. With the boom in advanced linguistic models, it is time to get rid of these limitations and to revolutionize the way in which we understand the feeling of customers, by discovering the context, emotion and intention that determines opinions.

Italian flags with wavy lines: the history of the analysis of feelings

The analysis of feelings has existed for a long time. Most companies use a version to analyze large volumes of text, whether publications on social networks, responses to surveys or comments on websites.

As a rule, each text extract is classified as “positive”, “negative” or “neutral”. The results are often displayed using a horizontal graph with green, white and red stripes. They represent the proportions of positive, neutral and negative comments, leading to the classic display of the feeling of “Italian flag”. There is a good chance that your organization has already used similar screens. They can be interesting, but their value is often limited.

Another current approach is to trace the percentages of positive and negative feelings over time, which gives rise to two corrugated lines which, although visually attractive, offer little usable information. At the beginning, some companies even tried to create a “feeling of feeling” by subtracting the percentage of negative comments from positive comments – yet another wavy line and yet another limited use measure.

Despite these faults, the analysis of the feelings was not completely ineffective. Its true potential lies in overcoming simple positive and negative percentages, often too volatile to be significant.

The categories of feelings can serve as a starting point for more in -depth analysis. For example:

What subjects are talking about people with positive feelings? How do they differ from the subjects mentioned negatively?

Visualization techniques such as butterfly graphics, which contrast the subjects mentioned positively on the one hand and negatively on the other, can highlight usable information to improve the customer experience.

It probably seems familiar. For more than 15 years, the promise to extract valuable marketing information from open comments has been attractive, but its practical value remains limited.

The good news is that recent progress in linguistic models make it possible to revisit the analysis of feelings in a new and exciting way. Thanks to these modern tools, we can discover what determines the perception of customers with regard to our products and services.

The limits of the traditional analysis of feelings

Before exploring the new features offered by modern language models, it is essential to understand the limits of the analysis of traditional feelings. Without this understanding, we risk repeating the same problems with new technologies. Three main problems have limited the effectiveness of the conventional analysis of feelings:

Lack of context

The traditional analysis of feelings generally works with text extracts, often omitting to take into account the context necessary to interpret the meaning with precision. For example, take the comment: “You are incredible. »»

Without context, feeling can be positive or negative. The feeling is probably negative if the person has evaluated their satisfaction with your service at 1/5. If they noted you 5/5, it’s probably positive. The context is essential, but the traditional analysis of feelings rarely integrates it.

Ambiguity in the neutral category

One of the reasons for the volatility of the percentages of positive and negative feeling is the neutral category, which often combines two very different groups:

Really neutral comments. Comments where feeling cannot be categorized.

These groups must be treated separately, but traditional systems fail to distinguish them. Many suppliers avoid approaching this problem due to the complexity involved, leaving this ambiguity irresol.

Simplify human expressions to the extreme

Categorize text as positive or negative simplifies the complexities of human language to the extreme. A single comment can be used for several purposes or express contradictory emotions, which makes an insufficient binary approach. This limitation obliges to frequently use the neutral category, which further reduces the accuracy and the value of the analysis.

Understanding these limits is essential to build better systems that prevent the traps of traditional analysis of feelings.

Dig more deeply: How to increase market studies and obtain customer information thanks to AI

How modern linguistic models change the situation

Modern linguistic models can revolutionize the analysis of feelings, but only if we avoid reproducing the defects of traditional methods. Creating a system of analysis of feelings using a large modern language model (LLM) is surprisingly simple, but it always comes up against the same limits.

For example, consider the following prompt:

“Please classify the following declaration as positive, negative or neutral. Give me only the category of one word: “You are incredible. »»

When using Gemini 1.5, the answer is:

Although a LLM feeds this feeling analysis system, its results are as limited as traditional approaches. It is simply easier to implement.

To truly exploit this potential, we must recognize the limits of the traditional analysis of feelings and seek more advanced solutions. Here are some ways whose current models can take up these challenges:

Include the context: The analysis of feelings must integrate contextual information, such as satisfaction rates, previous interactions or other relevant data, rather than analyzing isolated text extracts.

Adopt a more sophisticated categorization scheme: Go beyond positive, negative and neutral basic labels to include categories such as frustration, admiration, gratitude or sarcasm, adapted to what matters most to improve customer experience.

Focus on goal, not just emotion: Analyze the intention or atmosphere behind the declarations, whether descriptive, sarcastic, uncertain or informative. Understanding these nuances can lead to further information.

A new wave of analysis of feelings is emerging, which goes beyond the exceeded visualizations of the “Italian flag”. By effectively using modern tools, we can guarantee that our analysis of feelings provides usable and relevant information for the company.

Dig more deeply: From feeling to empathy: understand what customers feel

The authors contributors are invited to create content for Martech and are chosen for their expertise and their contribution to the Martech community. Our contributors work under the supervision of drafting And contributions are verified for their quality and relevance for our readers. The opinions they express are theirs.