ERAS: Bridging the Gap Between Code and Emotion in Spanish Learning Environments

TL;DR

Researchers have developed the Educational Resources Assessment System (ERAS) to capture how students feel while learning to code. By building a first-of-its-kind Spanish corpus labeled with learning-specific emotions like frustration and engagement, the study paves the way for Intelligent Tutoring Systems (ITS) that don't just teach—they empathize.

Background: Beyond "Happy" or "Sad"

In the world of AI education, recognizing that a student is "angry" isn't as helpful as recognizing they are frustrated by a specific recursion exercise. Existing Sentiment Analysis (SA) tools are often general-purpose and English-centric. The authors of this paper identify a critical vacuum: the lack of Spanish-language resources that focus on learning-centered emotions, which are the true drivers of information retention and cognitive success.

Methodology: The ERAS Framework

The researchers developed a multi-stage pipeline to transform raw student opinions into a structured dataset for machine learning.

1. Data Collection (ERAS)

Students interact with programming learning objects (videos, text, images). After consuming the content, they provide a textual opinion and tag it with one of five emotions: Frustrated, Bored, Neutral, Excited, or Engagement.

2. The Sentiment Analysis Module

The authors compared several classical Machine Learning architectures. The process follows a standard but robust NLP pipeline:

• Pre-processing: Normalization, tokenization, and stemming of Spanish text.
• Weighting: Generating a TF-IDF matrix (Term Frequency-Inverse Document Frequency) to quantify the importance of specific pedagogical terms.
• Classification: Testing algorithms including SVM, KNN, and Naive Bayes.

Figure 1: The ERAS system loop—from student interaction to emotion registration.

Experimental Insights & SOTA Comparison

The study evaluated six different classifiers. Interestingly, the Bernoulli Naive Bayes (BNB) emerged as the most robust, particularly for handling text features.

However, the transition from general text (TASS) to the niche educational corpus was challenging. The accuracy dropped to 40.70% on the newly created dataset.

The "Positivity Bias" Problem

A key finding of the research was the extreme imbalance in student feedback. Students were overwhelmingly positive (754 vs. 97 negative comments).

Figure 2: The distribution reveals a heavy bias towards positive emotions like Engagement and Excitement.

The authors argue that this "neutral/excited" tendency is an obstacle for training robust AI. To build a truly intelligent system, we must "induce students to write negative comments" to capture the nuances of where learning breaks down.

Critical Analysis & Future Outlook

This work serves as a foundational step for Spanish-speaking Intelligent Learning Environments (ILE). While the current accuracy on the internal corpus is modest, the methodology proves that we can quantify subjective learning experiences.

• The SOTA Edge: Unlike general sentiment tools, this approach explicitly links emotions to specific learning formats (Video vs. Image-Text), allowing teachers to see exactly which resources cause frustration.
• Limitations: The corpus size remains small (851 entries), and the gender imbalance (45 male vs. 8 female) might introduce demographic bias in linguistic expression.
• The Next Leap: Integration with actual Intelligent Tutoring Systems will allow for real-time pedagogical adjustments—switching a video to a text format if the system detects the student is becoming "bored" or "frustrated."

Takeaway

The future of AI in the classroom isn't just about better content; it’s about better listening. By focusing on Spanish learning-centered emotions, ERAS provides the raw data needed to make digital tutors as perceptive as human ones.