Literal, Metphorical or Both? Detecting Metaphoricity in Isolated Adjective-Noun Phrases

Agnieszka Mykowiecka, Malgorzata Marciniak, Aleksander WawerNAACL 2018

TL;DR

This paper tackles the challenge of automatically identifying whether Polish adjective-noun phrases are used literally, metaphorically, or can be both, using neural networks and word embeddings. They demonstrate that simple word embeddings alone can achieve competitive results, even outperforming more complex methods that incorporate additional linguistic features.

What Problem Does This Paper Address?

Natural language is inherently ambiguous, and a significant part of this ambiguity stems from the metaphorical use of words and phrases. While humans easily understand when a phrase like "iron nerves" is figurative, machines struggle. This difficulty poses a problem for various Natural Language Processing (NLP) applications, such as information extraction, machine translation, and textual entailment, where misinterpreting a metaphor can lead to incorrect results. Existing research often focuses on distinguishing purely literal from purely metaphorical phrases, neglecting the common case where a phrase can have both interpretations depending on context. This paper aims to address this gap by developing a system that can classify isolated Polish adjective-noun phrases into three categories: literal (L), metaphorical (M), and those that can be both (B).

Methods

The core of the proposed method relies on neural networks, specifically dense layers and LSTMs, to classify adjective-noun phrases. The input to these networks consists of word embeddings representing the individual words in the phrase. The authors explore various neural network architectures, including different numbers of dense layers and vector sizes for the embeddings. They also investigate the impact of incorporating additional linguistic features, such as the domain of the adjective and the abstractness/concreteness of the noun, to see if they improve classification performance.

Datasets

The primary dataset used in this research is a newly curated collection of Polish adjective-noun phrases, divided into three classes: literal (L), metaphorical (M), and both (B). This dataset is an extension of a previous resource and was manually annotated. To compare their findings with existing work, the authors also evaluated their method on two English datasets: one used by Tsvetkov et al. (2014) and another by Gutierrez et al. (2016). These English datasets primarily consist of literal and metaphorical phrases, excluding the "both" category.

Key Insights and Ideas

  • Word Embeddings are Powerful: The study highlights that pre-trained word embeddings alone, without complex linguistic features, are surprisingly effective for metaphoricity detection. This suggests that the semantic relationships captured by embeddings are sufficient for distinguishing literal from metaphorical meanings in many cases.
  • The "Both" Category Matters: By explicitly including a "both" (B) category, the authors acknowledge a more nuanced linguistic reality. However, they find that classifying these "both" phrases is significantly harder, with lower F1 scores compared to distinguishing between purely literal and purely metaphorical phrases.
  • Additional Features Don't Always Help: Counterintuitively, adding features like adjective domains or noun abstractness did not consistently improve the classification results. This reinforces the power of well-trained word embeddings and suggests that current methods for incorporating these features might not be optimal.
  • Architecture Simplicity Can Be Effective: The research shows that simpler neural network architectures (e.g., fewer dense layers, smaller vector sizes) can achieve performance comparable to more complex ones, especially when using only word embeddings.

Key Results

  • For Polish data, the best F1 scores for classifying into L, M, and B categories reached around 0.77-0.79.
  • When classifying only into L and M (excluding B), the accuracy improved significantly, reaching up to 0.97 on the English dataset from Gutierrez et al. (2016).
  • On the English dataset from Tsvetkov et al. (2014), their method achieved an accuracy of 0.84, slightly lower than the original paper's 0.86, but with better recall for metaphorical phrases.
  • The "B" category phrases proved to be the most challenging, with the best F1 result for this specific class being around 0.49.

Limitations and Future Work

The authors acknowledge that classifying phrases into L, M, and B categories in isolation is challenging and may not be fully reliable, especially for less frequent phrases. They suggest that the obtained results could serve as an additional source of information for such cases.

Future work planned by the authors includes focusing on phrases that exhibit both literal and metaphorical usages (the "B" category) and recognizing their specific usage at the sentence level. This indicates a move towards more context-aware metaphor detection.

Summary updated

BibTeX
@inproceedings{mykowiecka2018literal,
  author = {Agnieszka Mykowiecka and Malgorzata Marciniak and Aleksander Wawer},
  title = {{Literal, Metphorical or Both? Detecting Metaphoricity in Isolated Adjective-Noun Phrases}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2018},
}
Figure