MUESCA: A Multi-Genre Corpus and Annotation Scheme for Subjectivity in Spanish

In the vast landscape of human language, there is a constant tension between what is simply true and what is felt to be true. For decades, scientists working in the field of natural language processing have tried to teach computers to tell the difference between a statement of fact and an expression of personal opinion. This task, known as subjectivity detection, is crucial for everything from sorting news articles to understanding public sentiment. However, the traditional approach has often been too simple, treating language as a binary switch where a sentence is either objective or subjective. In reality, human communication is rarely so black and white. A news report might state a fact while subtly framing it with a journalist's perspective; a scientific paper might present hard data while using cautious language to describe future possibilities. The challenge lies in capturing this fluid middle ground, where the "author's fingerprint" is hidden inside words that appear neutral on the surface.

A team of researchers at the University of Alicante in Spain has taken a significant step toward solving this puzzle by creating a new resource called MUESCA. Rather than forcing texts into rigid boxes, they built a dataset that treats subjectivity as a sliding scale, or a continuum. To do this, they gathered and analyzed three very different types of writing: scientific abstracts, which represent the most objective end of the spectrum; consumer reviews, which represent the most subjective end; and news articles, which sit in the messy, hybrid middle. By examining how authors in these different genres use specific linguistic tools—such as words that show certainty, phrases that show doubt, or pronouns that reveal personal involvement—the researchers mapped out exactly how the "voice" of the writer changes across these styles. Their work provides a new way to measure the invisible presence of a human author within a text, offering a more nuanced map for computers to navigate the complex world of human expression.

The core of this project is the MUESCA corpus, a collection of 1,865 sentences drawn from 135 documents. The researchers did not just collect these sentences; they meticulously annotated them, marking specific grammatical and lexical features that signal how much the author is committing to what they are saying. They divided these signals into two main groups: high author commitment and low author commitment. High commitment includes markers of certainty, such as using the future tense or stating something as a fact, and markers of participation, like using "I" or "we." Low commitment includes markers of doubt, such as the subjunctive mood or words like "maybe," and markers of distancing, such as using passive voice to hide who did the action. By counting these features, the team could see how scientific abstracts tend to rely on specific data and impersonal structures, while reviews are filled with personal opinions and direct engagement. News articles, as expected, showed a mix of both, confirming their status as a hybrid genre.

To ensure their findings were robust, the researchers tested whether these linguistic markers could actually help computers distinguish between the different types of writing. They trained machine learning models to classify the texts into objective, hybrid, or subjective categories. When the models were given only the raw text, they performed reasonably well, but when the researchers added the specific pragmatic markers from their annotation scheme, the models became significantly better at the task. The most effective combination of features involved measuring the level of specificity and the degree of certainty. This approach allowed the computer to understand not just what the text was about, but how the author was presenting that information. The results were particularly strong for the news genre, which is often the hardest to classify because it blends facts with interpretation. By using these pragmatic markers, the model improved its ability to spot the subtle ways journalists frame their stories.

The researchers also explored how these human-designed rules could work alongside modern artificial intelligence. They combined their specific markers with a powerful language model called BETO, which is trained on vast amounts of Spanish text. Even though this advanced model was already very good at understanding language, adding the specific subjectivity markers still improved its performance. The improvement was small in percentage terms, but statistically significant, suggesting that the explicit rules about author commitment helped the computer resolve ambiguities that pure statistical patterns missed. This is especially true for news articles, where the model's accuracy increased when it could see the specific linguistic signals of doubt or certainty. The study confirms that while large AI models are powerful, they still benefit from a structured understanding of how humans construct their arguments and express their beliefs.

One of the most revealing aspects of the study was how the news genre behaved when pitted against the other two extremes. When the researchers tested the model by asking it to distinguish news from scientific abstracts, the model found it relatively easy to tell them apart. However, when asked to distinguish news from consumer reviews, the task was harder. This suggests that the language of journalism, despite its reputation for neutrality, shares more pragmatic DNA with personal opinion pieces than with the detached language of science. The news articles often contained enough subjective framing and evaluative language to blur the line with reviews, making them harder to separate from the subjective pole. This finding supports the idea that subjectivity is not a fixed property of a text type but a fluid quality that can shift depending on how the author chooses to present their information.

The MUESCA project does not claim to have solved the problem of subjectivity detection entirely, but it provides a solid foundation for future work. The researchers have made their dataset and their annotation guidelines available to others, allowing the scientific community to build upon their work. They plan to expand the corpus to include more genres, such as social media posts and technical reports, to see if their method holds up in more informal or polarized environments. They also see potential applications in detecting deception, where the way information is presented—through certainty or doubt—might be more telling than the facts themselves. By focusing on the pragmatic delivery of information rather than just the content, this approach offers a new lens for understanding how humans use language to shape reality. The work demonstrates that by breaking down the complex task of subjectivity into measurable, linguistic components, we can build tools that are not only more accurate but also more transparent about how they reach their conclusions.