Why Analyzing Text Feedback Is Harder Than It Looks

Meet Maya, a brand manager at a popular beverage brand, perhaps your favourite one. She just launched a new packaging design for the brand and is now reviewing feedback from surveys and social media.

She knows this feedback matters, every comment is a signal. But making sense of it? That’s the hard part.

Most companies collect open-text feedback but underestimate how messy it gets when you try to extract patterns from it.

Unlike rating scales, open answers come in unpredictable forms. Some people write short phrases. Others send paragraphs filled with emotion, sarcasm, or even conflicting tones in one sentence. This makes unstructured data extremely difficult to process at scale.

According to IBM, nearly 80% of enterprise data is unstructured. That includes emails, survey responses, social posts, chats, and more. Yet most businesses rely on tools designed for structured input (like rows and columns) to make sense of it.

One way for users to acquire patterns from unstructured data is by relying on the methodology called sentiment analysis. Sentiment analysis itself is a natural language processing (NLP) technique used to determine the emotional tone behind a body of text. The main goal is to identify whether the expressed opinion in a piece of text is positive, negative, or neutral.

Sentiment analysis is a common method for understanding the intent behind text, but it is not as simple as labeling it positive or negative based on the presence of certain words. While we can subjectively assign sentiment to individual texts, what happens when the data consists of hundreds, thousands, or even millions of entries? This is where machine learning models can help us automatically predict sentiment.

At ARIF, we use machine learning models specifically trained to predict text sentiment in both Indonesian and English. By leveraging the power of data and AI technology, we ensure our results are validated and allow users to further explore and discuss insights from sentiment analysis.

However, even the best models can misread context. “This product is sick” might be flagged as negative by older lexicon-based tools, but it could be positive comment depending on the context. And sarcasm? It’s still one of the hardest things for machines to detect (Pang & Lee, 2008).

Another issue is noise. In one internal review, 15% of feedback lines were unusable due to typos, blank fields, or mixed languages. Without preprocessing, even the best analysis tool gives you garbage-in-garbage-out results.

To fix this, ARIF add an option for users to define custom stopwords, which are words that add no analytical value and differ by context. By removing these unhelpful words during the cleaning stage, ARIF helps improve data quality and clarity before analysis. This is especially helpful if users already know which words are not helpful in the process of gaining insight.

And don’t forget about annotation bias. A 2021 study by Hovy & Prabhumoye found that even human raters often disagree on sentiment labels. So if your baseline isn’t clear, your model won’t be either.

To bring more clarity, ARIF use a model to determine sentiment confidence scores for each piece of feedback. These scores help users gauge how strongly a response leans positive or negative, with no manual guesswork needed. By combining AI technology, we can confidently understand the overall sentiment from your textual data while also providing users with a way to engage in follow-up discussions based on the results.

Analyzing text feedback isn’t just a “click run sentiment tool” type of task. It takes structure, cleaning, the right model, and clear definitions of what you’re actually measuring.

Before jumping to charts and dashboards, ask yourself:

• Are the responses clean enough to process?
• Is the model trained in my industry language?
• Does my team know how to validate the output?

Unstructured data can unlock deeper signals, but only when treated with precision.

One of ARIF’s key features for providing deeper insights is a comprehensive report on sentiment analysis from your text data. ARIF not only classifies text sentiment but also leverages large language models (LLMs) to interpret overall patterns. The LLM analyzes word cloud outputs and incorporates prior user prompts to deliver contextual insights on key themes.

If you’re working with large volumes of survey or customer feedback and need to simplify the analysis process, ARIF can help. It makes sentiment analysis easier to apply in real business settings without deep technical setup.