TL;DR
This paper tackles the challenge of reliably detecting emotions in real-world call center conversations, which are notoriously complex and context-dependent. By comparing two distinct French call center corpora, they demonstrate that while cross-corpus emotion detection is difficult, it's not impossible, offering insights into how context influences emotion expression and detection.
What Problem Does This Paper Address?
Building emotion-aware spoken dialog systems is a key goal for more natural human-computer interaction. However, most emotion detection research relies on artificially induced or recalled emotions, which lack the ecological validity of real-life interactions. Collecting and annotating naturalistic emotional data is expensive and fraught with privacy concerns, leading to a scarcity of such datasets. This paper specifically addresses the challenge of detecting emotions in real-life call center environments, which are rich in spontaneous emotional expression but also highly context-dependent. The authors aim to understand if models trained on one type of call center data can generalize to another, and how different contexts (medical emergency vs. service complaints) affect emotion expression and detection.
Methods
The researchers employed a cross-corpora evaluation strategy. They trained emotion detection models on one corpus and tested them on another to assess generalization capabilities. For acoustic feature extraction, they utilized the openEAR platform, a standard tool in the field, to ensure comparability with existing research. They experimented with two feature sets: a reduced set (384 features) and a baseline set (988 features). For classification, they used Support Vector Machines (SVMs) with the libSVM implementation. To enable cross-corpus comparison, they focused on three macro-classes of emotions: Anger, Positive, and Neutral, and ensured balanced classes for training.
Datasets
The study utilized two French call center corpora:
- CEMO corpus: Collected from a medical emergency call center, featuring a wide range of emotions including fear, stress, panic, and anger, often expressed by callers seeking medical advice.
- EmoVox corpus: Collected from a service complaints call center (EDF, French electricity utility), focusing on emotions related to customer service interactions, such as annoyance and satisfaction.
Both corpora contain naturally occurring spoken data, with manual annotations for emotional states.
Key Insights and Ideas
- Context is King (and Complicates Things): The study highlights that emotion expression, particularly for positive states, varies significantly between different call center contexts (medical emergency vs. service complaints). This makes direct cross-corpus generalization challenging.
- Anger is More Universal (Relatively): Anger-related states appear to be more transferable across different call center contexts compared to positive emotions. Models trained on anger in one corpus showed somewhat better, though still limited, performance when tested on the other.
- Agent vs. Client Differences: The paper notes distinct patterns in emotion expression between agents and clients. Clients tend to express a wider range of emotions, including more intense anger, while agents' expressions are more controlled and task-oriented.
- Annotation Scheme Matters: The authors acknowledge that differences in annotation schemes (even when using macro-classes) and the number of annotators can impact comparability and generalization.
Key Results
- Cross-corpus evaluation yielded generally low performance, indicating that models trained on one call center context do not directly transfer well to another.
- For instance, when training on CEMO's "Anger1" data, the F-score for Anger was 0.59, but almost no instances were classified as Positive. Conversely, training on EmoVox improved Positive classification but worsened Anger detection.
- The use of 384 features versus 988 features did not significantly impact performance in these cross-corpus experiments, suggesting that a more concise feature set might be sufficient for future work.
- The study found that Anger classes were more similar between corpora than Positive classes.
Limitations and Future Work
- Annotation Scheme Incompatibility: The primary limitation acknowledged is the non-exact compatibility of the annotation schemes between the two corpora, making direct comparisons difficult.
- Contextual Nuances: The authors recognize the need to better understand the factors influencing generalization power, particularly the influence of specific contexts.
- Future Work: The researchers plan to:
- Investigate the "affective disposition taken towards another person" in specific communication contexts.
- Develop models that are robust to the detection of emotions like anger in real-life interactions.
- Explore the use of blended corpora with a common annotation system to improve generalization.
- Conduct future work focusing on specific models for Agent and Client tasks to better compare their expressed emotions.
Summary updated