Text-as-Signal: Quantitative Semantic Scoring with Embeddings, Logprobs, and Noise Reduction

Researcher Hugo Moreira has introduced 'Text-as-Signal,' a novel methodological pipeline for transforming unstructured text into structured, quantitative semantic signals. The framework employs a three-stage process: first, generating full-document embeddings (using models like Qwen); second, scoring each document via logprob-based evaluation against a user-defined 'positional dictionary' of semantic concepts; and third, projecting the results onto a noise-reduced, low-dimensional manifold using techniques like UMAP for clearer structural interpretation. This approach moves beyond simple topic modeling to provide a configurable, quantitative score for how a document relates to specific semantic dimensions.

In a practical demonstration, Moreira applied the pipeline to a corpus of 11,922 Portuguese-language news articles about Artificial Intelligence. The configurable dictionary was instantiated as six distinct semantic dimensions, allowing each article to be positioned within this 'identity space.' The resulting data supports both granular document-level analysis and high-level corpus characterization through aggregated profiles. The paper details how the integration of embeddings, semantic indicators from the model's output space, and a three-stage anomaly-detection procedure creates an operational workflow for real-world AI engineering tasks, including corpus inspection, trend monitoring, and providing support for downstream analytical applications.

The key innovation is the framework's adaptability. Because the core 'identity layer'—the dictionary of semantic dimensions—is configurable, the same technical pipeline can be repurposed for different analytical needs across various domains, rather than being locked into a single, universal schema. This makes it a powerful tool for researchers and engineers who need to systematically measure and track semantic content over time within large document collections.