TL;DR
This paper investigates why we choose metaphorical language over literal alternatives in specific contexts. By building and testing five models based on linguistic and cognitive features, the authors found that common discourse properties like abstractness, affect, and coherence don't reliably predict metaphorical usage, suggesting a need for more nuanced approaches in NLP.
What Problem Does This Paper Address?
Metaphors are fundamental to human communication, but understanding why they are used in specific instances, rather than their literal counterparts, remains a challenge. While Natural Language Processing (NLP) has made strides in detecting and interpreting metaphors, the driving forces behind their selection in discourse are less understood. This paper aims to bridge this gap by exploring whether established linguistic and cognitive features of discourse can predict the choice between a metaphorical and a synonymous literal expression. This is crucial for developing more human-like language understanding and generation systems in NLP.
Methods
The authors developed five distinct models to predict the use of metaphorical vs. literal expressions. These models are based on different discourse properties:
- Frequency Approach: A baseline model that predicts the most frequent expression (metaphorical or literal) based on its occurrence in the ukWaC corpus.
- Cognitively-Inspired Approaches:
- Abstractness (Abstr.): Models abstractness of the preceding discourse using word abstractness scores from Brysbaert et al. (2014).
- Emotionality (Emo.): Models the emotional load of the preceding discourse using an English emotion lexicon from Buechel et al. (2020).
- Discourse Coherence Approach (LCG): Adapts a Lexical Coherence Graph to measure semantic relatedness between words in the preceding context and the target expression using contextualized BERT embeddings.
- Contextualized Word Representations (BERT): Utilizes pre-trained language models (BERTbase and BERTlarge) in a zero-shot cloze-task manner to predict the masked expression.
Datasets
The study utilizes a recently released dataset specifically designed for this task, introduced by Piccirilli and Schulte im Walde (2021, 2022). This dataset contains:
- 1,000 discourses of five to six sentences each, extracted from the ukWaC corpus.
- Each discourse's final sentence features either a metaphorical expression or its literal alternative from synonymous subject-verb (SV) or verb-object (VO) pairs.
- These pairs were selected from existing resources (Shutova, 2010; Mohammad et al., 2016).
- Crucially, the dataset includes crowdsourced human judgments on which expression (metaphorical or literal) annotators favored in the context of 287 discourses where at least 70% of annotators agreed.
Key Insights and Ideas
- No Single Driver: The study's core finding is that none of the tested discourse properties (frequency, abstractness, affect, coherence) individually or collectively seem to be sufficient to systematically explain why a metaphorical expression is chosen over a literal one.
- Frequency is Not King (for choice): While frequency influences language use, it doesn't fully explain the preference for metaphorical vs. literal expressions. Some less frequent metaphors are preferred by humans, and some more frequent ones are not.
- Abstractness and Affect Don't Guarantee Metaphor: Models based on abstractness and emotionality of the discourse did not consistently align with human preferences, suggesting these factors aren't direct triggers for metaphorical choices.
- Coherence is Complex: Lexical coherence, while important for discourse structure, didn't strongly predict metaphorical preference, challenging the idea that discourse context directly dictates metaphorical choice.
- PLMs Show Frequency Bias: The strong overlap between BERT predictions and the frequency baseline suggests that large language models might be implicitly learning frequency patterns rather than deeper semantic or cognitive drivers of metaphor choice.
- Human Preferences are Nuanced: The analysis of human annotations reveals complex patterns that are not fully captured by the tested computational models.
Key Results
- All five models achieved around 50% accuracy when predicting metaphorical vs. literal expressions against the original, balanced ukWaC data. However, their individual prediction patterns varied significantly.
- The frequency baseline predicted the literal expression 74% of the time, aligning with the observation that literal usages were more frequent in the corpus for most pairs.
- Abstractness and emotionality models tended to predict metaphorical usages more often (up to 100% for Emo.), suggesting they might be overestimating the influence of these factors.
- BERT models predominantly predicted literal expressions (around 80%), indicating a potential bias towards more common or literal language.
- When comparing model predictions to human annotations (Figure 6), the plots show that while some expressions fall on the regression line (indicating agreement), many points deviate, highlighting the models' inability to fully capture human preferences. For instance, some expressions predicted as literal by models were strongly favored as metaphorical by humans.
Limitations and Future Work
- Context Window Size: The study used discourses of five to six sentences. Future work could explore how different context window sizes impact findings.
- PLM Attention Mechanisms: Investigating BERT's attention mechanisms and fine-tuning these models could potentially improve their ability to mimic human preferences.
- Exploring More Features: The authors suggest exploring additional discourse properties like co-reference, complexity, aptness, creativity, genre, and specific domains to gain a more comprehensive understanding.
- Nuanced Approach Needed: The findings strongly suggest that a more nuanced approach is required to understand and model metaphorical language choices in NLP.
Summary updated