Guidelines for the Annotation of Deliberate Linguistic Metaphor

Stefanie Dipper, Adam Roussel, Alexandra Wiemann, Won Kim, Tra-My NguyenNAACL 2024

TL;DR

This paper introduces a new set of guidelines for annotating "deliberate linguistic metaphors" – those intentionally used by speakers to convey meaning. By focusing on these deliberate instances, the authors have created a corpus of TEDx talks that reveals a significantly higher proportion of such metaphors than previously identified, offering a richer understanding of how speakers use language creatively.

What Problem Does This Paper Address?

Understanding how language is used metaphorically is crucial for fields like natural language processing and linguistics. However, previous annotation efforts often focused on all types of metaphors, including conventionalized ones that speakers might not even be aware of. This broad approach diluted the focus on deliberate metaphors – those that are intentionally crafted and often signal a speaker's argumentative moves or creative intent. The challenge, therefore, was to develop clear guidelines and a methodology to systematically identify and annotate these deliberate metaphors, moving beyond simply flagging any word that deviates from its literal meaning.

Methods

The core contribution of this paper is the development of annotation guidelines specifically for deliberate linguistic metaphors. These guidelines aim to identify expressions that are either novel or used in an unusual way for a known metaphorical image. They also consider metaphors that are "marked" in some way, such as through unusual grammatical constructions, typographical emphasis, or appearing in the title.

The annotation process involves identifying "metaphor-related words" (MRWs) and then classifying them. A key aspect is the concept of "MRW chains," where multiple MRWs contributing to the same metaphorical image are linked together. Within these chains, a central "anchor" expression is identified, which grounds the metaphorical image in reality. The guidelines also introduce labels like "deliberate," "extended" (for metaphors spanning multiple sentences), "grey area" (for metaphors with characteristics of both deliberate and conventionalized types), and "revitalized" (for conventionalized metaphors used in a new light).

Datasets

The authors created a new annotated corpus of ten German TEDx talks, totaling approximately 20,000 tokens. Four of these talks were doubly annotated by two of the authors to assess inter-annotator agreement. This corpus is made freely available. The authors also reference and build upon existing corpora and annotation procedures, such as the VUAMC (VU Amsterdam Metaphor Corpus) and the MIPVU guidelines, but differentiate their work by focusing specifically on deliberate metaphors.

Key Insights and Ideas

  • Focusing on Deliberate Metaphor Yields Higher Proportions: By specifically targeting deliberate metaphors, the authors found that 1.35% of tokens in their TEDx corpus were classified as such. This is significantly higher than the estimated proportions of deliberate metaphors in previous studies, suggesting that previous methods may have overlooked many instances.
  • The Importance of "Markedness": The guidelines emphasize that deliberate metaphors are often "marked" in some way. This can include novel usage, unusual grammatical constructions, or even typographical emphasis, all of which signal the speaker's intentionality.
  • MRW Chains Capture Metaphorical Complexity: The concept of MRW chains allows for the annotation of complex metaphorical expressions that extend beyond single words, capturing how multiple linguistic elements contribute to a single metaphorical image.
  • Anchor as a Grounding Element: The "anchor" concept is crucial for identifying the literal referent of a metaphor, helping to clarify the cross-domain mapping and the intended meaning.

Key Results

The primary result is the creation of a new annotated corpus and the associated guidelines for identifying deliberate linguistic metaphors. The corpus analysis revealed that 1.35% of tokens in the TEDx talks were classified as deliberate metaphorical expressions. The paper also reports inter-annotator agreement scores (using the gamma agreement measure) ranging from 0.35 to 0.56 for the evaluation texts, which are considered promising given the complexity of the task and annotation scheme. Table 1 shows the distribution of MRW types in their corpus: 85 "deliberate" chains (264 tokens), 25 "extended" chains (52 tokens), 15 "grey area" chains (30 tokens), and 20 "revitalized" chains (46 tokens).

Limitations and Future Work

The authors acknowledge that the increased complexity of their annotation scheme, while capturing more nuance, also makes evaluating reliability more challenging. They also note that the current corpus consists solely of TEDx talks, and it's unclear how much the text type contributes to the high proportion of deliberate metaphors. Future work will involve expanding the corpus to include other text genres to investigate this further. They also suggest that further refinement of the annotation guidelines could improve inter-annotator agreement.

Summary updated

BibTeX
@inproceedings{dipper2024guidelines,
  author = {Stefanie Dipper and Adam Roussel and Alexandra Wiemann and Won Kim and Tra-My Nguyen},
  title = {{Guidelines for the Annotation of Deliberate Linguistic Metaphor}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2024},
}
Figure