情感计算在现实生活中的应用受到限制，因为需要使用特殊设备，如脑电图 和心电图，这些设备对于日常使用来说不实际。本研究使用一种用户友好的耳机 式温度计引入了新的情感识别数据——鼓膜温度(TMT)。虽然 TMT 在之前的 研究中被用作脑活动偏侧化的指标，但其在情感识别方面的潜力尚未被完全探索。 当前研究包括三个自我实验和三个群体实验。前两个自我实验在自然环境中进行， 显示出左右 TMT 绝对差异与负面情绪之间的正相关性，与先前研究的结果一致。 在最后一个实验中，采用了线性 Support Vector 分类器来利用左右 TMT 差异，其 分类准确度高于随机水平。三个群体实验采用了典型的情绪诱导方法:观看短视 频，回忆自传经历和想象情景。虽然在这些实验中没有找到 TMT 与情感之间的一 致相关性，但使用左右 TMT 差异的四种特征的非线性 Support Vector 分类器在所 有实验中的情绪分类，以及在两个实验中的积极情绪与负面情绪分类中实现了高 于随机水平的准确度。值得注意的是，这种分类准确度的提高仅在使用左右 TMT 差异时观察到，并未在仅左耳或右耳 TMT 或手腕皮肤温度中发现。这些发现表明 我们的模型可以从左右 TMT 差异中提取独特的信息，使 TMT 在区分不同情绪方 面具有价值，并适用于各种情境。由于使用耳机测量的便捷性，TMT 有潜力促进 情感识别在现实生活中的实际应用。
The real-life application of affective computing is limited due to the requirement of specialized devices such as electroencephalograms and electrocardiograms, which are impractical for daily usage. This study introduces new data for emotion recognition—tympanic membrane temperature (TMT)— using a user-friendly, earphone-type thermometer. While TMT has been utilized as an index for lateralized brain activity in previous research, its potential for emotion recognition has not been fully explored. The current study consists of three self-experiments and three group experiments. The first two self- experiments were conducted in naturalistic settings, revealing a positive correlation between absolute right-left TMT and negative emotion, consistent with the findings in previous studies. The last self-experiment was conducted in an experimental environment, and a linear Support Vector Classifier using right- left TMT achieved classification accuracy of discrete emotion higher than the chance level. Subsequently, three group experiments was conducted using three typical emotion induction methods: short video watching, autobiographical recalling, and scenario imagination. Although a consistent correlation between TMT and emotion was not found in the group experiments, a nonlinear Support Vector Classifier using four features of right-left TMT achieved higher-than- chance-level accuracies in discrete emotion classification across all experiments, as well as in positive versus negative emotions classification in two experiments. Notably, this improvement in the accuracy was observed exclusively in right-left TMT and not in the right, left TMT alone or the right, left and right-left wrist skin temperature. These findings indicate that our model can extract unique information from right-left TMT, making TMT valuable for distinguishing different emotions across various settings. With the ease of measurement using earphones, TMT can promote the practical application of emotion recognition in real-life scenarios.
