Introduction
As the global population ages, the preservation of "cognitive reserve"—the brain's capacity to improvise and solve problems through alternative neural pathways—has become a critical area of research. Mental exercise is a primary driver of this reserve. Psycholinguistic studies indicate that processing novel metaphors (e.g., "acoustic doodlings") requires significantly more cognitive effort than processing conventional metaphors (e.g., "the leg of a table"). While patients with Alzheimer’s often retain the ability to understand conventional metaphors, they frequently struggle with novel ones, making metaphor discussion a promising therapeutic tool.
However, current companion robots are limited by simple interfaces and canned speech. Furthermore, most Natural Language Processing (NLP) research focuses on binary metaphor detection rather than quantifying the "novelty" or creativity of a metaphor. This paper addresses these gaps by introducing a new dataset and a deep learning model designed to score metaphor novelty, providing a foundation for cognitively stimulating human-robot dialogue.
The Metaphor Novelty Dataset
To facilitate this research, a substantial new benchmark was developed: the Metaphor Novelty Dataset. This corpus consists of 18,439 syntactically-related word pairs. Each pair is scored on a scale from 0 (non-metaphoric) to 3 (highly novel).
To ensure data quality despite the subjectivity of metaphor interpretation, the researchers used a regression-based approach to aggregate crowdsourced labels from Amazon Mechanical Turk. This method effectively filtered "noise" and human bias more reliably than simple averaging. For validation, a subset of 3,162 instances was labeled by experts to serve as a high-quality test set.
Methodology and Novelty Scoring
The proposed system utilizes a deep neural network trained to predict novelty scores. The model incorporates traditional metaphor detection features alongside new features specifically tailored for novelty. A key insight of this work is that metaphor novelty is a distinct linguistic dimension; identifying "brain-teasers" requires more than just detecting the presence of a metaphor.
Beyond scoring, the research proposes a dialogue generation strategy called "Questioning the Author" (QtA). This pedagogical approach uses introspective questions—such as "What do you think the author wants us to know by metaphorically saying [novel metaphor]?"—to prompt deeper thinking and cognitive engagement in the user.
Performance Evaluation
The deep learning approach established a strong performance benchmark for metaphor novelty scoring, significantly outperforming several baselines.
As shown in the results, the model achieved a Pearson correlation ($r$) of 0.4600 and a Root Mean Square Error (RMSE) of 0.6502. Notably, when compared against a modified state-of-the-art metaphor detection system (Tsvetkov et al., 2014), the proposed model performed significantly better ($r = 0.46$ vs $r = 0.27$). This disparity confirms that novelty scoring is a distinct and more complex task than simple binary detection.
Conclusion and Future Work
This research demonstrates that robots can serve as unique social catalysts for cognitive therapy, providing a non-judgmental, embodied presence that facilitates introspective questioning. While the scoring model is complete, future work will focus on the full integration of this model into a physical robot and the automation of the QtA question-generation system. Subsequent user studies with elderly participants will be necessary to prove the clinical effectiveness of this system as a tool for cognitive exercise.
Summary updated