Abstract
Detecting when an individual is hiding critical information they possess—concealed information—is a distinct challenge from identifying outright lies. This study introduces a multi-task learning framework designed to spot these subtle behavioral cues in both text and speech. By analyzing a unique corpus of professional wine tasters engaged in a "blind tasting game," the researchers identified specific acoustic and linguistic "tells" that betray hidden knowledge. The resulting model significantly outperforms human domain experts, achieving an F1-score of 71.51% compared to the human average of 56.28%.
Introduction: Deception vs. Concealment
While deception detection (identifying false statements) is a well-established field, "concealed information detection" remains less explored. The distinction lies in the relationship between truth and appearance. In a deceptive scenario, a speaker lacks information but pretends to have it. Conversely, in concealment, the speaker possesses information but pretends they do not.
This behavior is particularly difficult to detect because concealers often experience lower cognitive load than traditional liars. Since they actually know the truth, they may appear more confident and consistent. This study seeks to identify the subtle signals that distinguish "honest expertise" from "unfairly obtained knowledge," such as in cases of corporate leaks or insider trading.
Methodology
The researchers developed a pipeline that integrates multiple data modalities through a Multi-task Learning (MTL) framework.
Feature Extraction
The model processes three primary types of features:
- Acoustic-Prosodic: Low-level descriptors including pitch, intensity, and jitter, alongside 13 Mel-Frequency Cepstral Coefficients (MFCCs).
- Linguistic: Semantic classes derived via LIWC (Linguistic Inquiry and Word Count), n-grams, and GloVe word embeddings.
- Individual Metadata: Contextual features such as gender, native language, and professional credentials.
Model Architecture
The core architecture utilizes a Bidirectional LSTM (BiLSTM) for processing sequential text data and a Multi-Layer Perceptron (MLP) for acoustic features. The "multi-task" component is critical: the model does not just predict concealment; it simultaneously learns to predict the identity of the wine and the correctness of the speaker's guess. This auxiliary training helps the model contextualize the technical reasoning provided by the speaker.
The Blind Tasting Dataset
To train and validate the model, the researchers created the Blind Tasting Game dataset. In this setup, 41 wine professionals participated in 49 sessions. In each round, one participant was randomly "informed" of the wine's identity and tasked with describing it while concealing their prior knowledge. This resulted in 164 hours of audio, providing a ground-truth labeled dataset of 3,288 multi-turn segments.
Results and Analysis
The MTL model demonstrated a clear advantage over both traditional machine learning baselines and human experts.
Human vs. Machine Performance
Human experts (other wine professionals) struggled with the task, achieving an F1-score of only 56.28%. They frequently mistook the "confidence" of concealers for genuine expertise. In contrast, the full MTL model achieved an F1-score of 71.51% on multi-turn segments, representing a 15% improvement over human performance.
Behavioral "Tells"
The study revealed several key insights into the behavior of concealers:
- The Confidence Offset: Concealers spoke with higher pitch, intensity, and a faster rate. This suggests that possessing the truth provides a physical confidence boost, contrary to the increased cognitive load often seen in liars.
- Linguistic Markers: Truthful speakers used significantly more filler pauses (e.g., "uh," "um"), reflecting the genuine cognitive effort of blind identification. Concealers used more "certainty" words and showed higher disparity between their spoken descriptions and written notes.
Conclusion
This research demonstrates that concealed information leaves detectable traces in speech and text that often elude human observers. By leveraging multi-task learning and multi-modal features, the model successfully identified these subtle cues. Future work may explore phonotactic variations and the temporal dynamics of concealment detection, with potential applications ranging from forensic interviews to financial market oversight.
Summary updated