TL;DR
This paper introduces "Latent Affective Mapping" (LAM), a novel framework for analyzing emotions in text. LAM leverages two distinct semantic spaces – one for general domain knowledge and another for affective fabric – to better understand and classify emotions, outperforming traditional methods.
What Problem Does This Paper Address?
Generating expressive synthetic speech requires accurately identifying the emotion conveyed by input text. Traditional methods often rely on manually curated lists of "emotional keywords," which are limited in scope and can miss nuanced emotional expressions. This paper addresses the challenge of building a more robust and data-driven emotion detection system that can capture the underlying affective nuances of language without relying on explicit keyword lists.
Methods
The core of the paper is the "Latent Affective Mapping" (LAM) framework, which builds upon Latent Semantic Analysis (LSA). LAM operates in two phases:
- Domain Corpus Analysis: A standard LSA is performed on a general domain corpus (e.g., English sentences for TTS) to create a "domain space." This space captures the topical structure of the language.
- Affective Corpus Analysis: A separate LSA is performed on a corpus of mood-annotated texts (e.g., blog entries) to create an "affective space." This space captures the affective fabric of the language.
The novelty lies in how these two spaces are integrated:
- Latent Affective Folding: Affective anchors (representations of emotions) from the affective space are "folded" into the domain space. This method is sensitive to the distribution of words in the affective corpus.
- Latent Affective Embedding: This is a more advanced technique. It involves mapping common words between the two corpora into a shared space, then transforming the affective anchors into the domain space. This embedding process aims to be less sensitive to corpus-specific word distributions and provides a richer representation of affective anchors.
Emotion classification is then performed by comparing the representation of an input text in the domain space to these embedded affective anchors using a cosine similarity measure.
Datasets
- Domain Corpus: Approximately 8,500 English sentences compiled for a concatenative text-to-speech voice.
- Affective Corpus: Approximately 5,000 mood-annotated blog entries from LiveJournal.com. These moods were mapped to 6 emotions (ANGER, DISGUST, FEAR, JOY, SADNESS, and SURPRISE).
- Evaluation Dataset: The "Affective Text" portion of the SemEval-2007 corpus, consisting of 1,250 short news headlines annotated with the same 6 emotions.
Key Insights and Ideas
- Two Semantic Spaces are Better Than One: The paper convincingly argues that separating domain knowledge from affective fabric and then finding a way to connect them leads to better emotion detection.
- Data-Driven Emotion Detection: LAM moves away from brittle keyword-based approaches towards a more flexible, data-driven method that learns emotional patterns from corpora.
- Latent Affective Embedding's Advantage: The "embedding" method shows superior performance, suggesting that a more sophisticated mapping between affective and domain spaces can better handle subtle emotional distinctions and corpus biases.
- Implications for Expressive TTS: The success of this framework directly benefits the generation of naturally expressive synthetic speech by providing a more accurate understanding of the intended emotion in text.
Key Results
The paper reports experimental evaluations on the SemEval-2007 "Affective Text" dataset. Both LAM instantiations (folding and embedding) significantly outperformed baseline methods:
- Latent Affective Folding: Achieved an F-measure of 31.1%.
- Latent Affective Embedding: Achieved an F-measure of 34.0%.
These results are substantially higher than those obtained by simple word accumulation (4.6% F-measure) and various LSA-based approaches using WordNet synsets (ranging from 19.6% to 21.1% F-measure). The latent affective embedding method, in particular, showed the best performance, indicating its effectiveness in capturing nuanced emotional information.
Limitations and Future Work
- Corpus Sensitivity: While latent affective embedding is less sensitive than folding, the performance of LAM is still dependent on the quality and representativeness of the training corpora.
- Future Work: The authors suggest that this framework is a promising solution for automatic emotion analysis, paving the way for its general deployability in applications like naturally expressive speech synthesis. Further exploration of different mapping instantiations and larger, more diverse datasets could yield even better results.
Summary updated