Automatic Metaphor Interpretation as a Paraphrasing Task

Ekaterina ShutovaNAACL 2010

TL;DR

This paper introduces a novel approach to automatically interpreting metaphors by reframing it as a paraphrasing task. Instead of relying on hand-crafted knowledge, their system learns to generate literal paraphrases for metaphorical expressions, achieving high accuracy and offering a more transferable solution for NLP applications.

What Problem Does This Paper Address?

Understanding metaphors is a significant challenge for Natural Language Processing (NLP) systems. Metaphors are pervasive in everyday language, allowing us to conceptualize abstract ideas in terms of more concrete ones (e.g., "killing a process" uses the domain of life and death to describe termination). This metaphorical usage creates a bottleneck for automatic text understanding, as literal interpretation methods fail. Existing approaches often depend on extensive, manually curated knowledge bases, which are difficult to scale and maintain. This paper aims to bridge this gap by developing a system that can automatically identify and interpret metaphors, producing literal paraphrases that can be readily integrated into other NLP applications.

Methods

The core of the system's approach is to treat metaphor interpretation as a paraphrasing task. Given a sentence with a metaphorical expression (specifically, a verb used metaphorically), the system performs the following steps:

  1. Paraphrase Generation: It generates a list of potential literal paraphrases for the metaphorical expression by leveraging a large corpus.
  2. Paraphrase Ranking: These paraphrases are then ranked based on their likelihood of co-occurring with the surrounding words in the original context. This is modeled as a joint probability calculation using maximum likelihood estimation.
  3. Shared Meaning Identification: To refine the paraphrases, the system uses WordNet to identify shared semantic features between the potential paraphrase and the original metaphorical term. It filters paraphrases that don't share common hypernyms or are not hypernyms themselves.
  4. Selectional Preference Filtering: The system then filters out irrelevant or still metaphorical paraphrases by employing automatically induced selectional preferences. It measures how well a verb's arguments fit its typical semantic constraints, using a measure derived from Resnik's work on selectional association. This helps identify truly literal interpretations.
  5. Sense Disambiguation (WSD): After identifying literal paraphrases, the system can also perform Word Sense Disambiguation by mapping the literal paraphrase to WordNet nodes that share common hypernyms with the original metaphorical verb.

Datasets

  • British National Corpus (BNC): A subset of the BNC was used for both training and evaluation. This subset contained 761 sentences and 13,642 words, sampled from various genres including literature, news, politics, international relations, history, and transcribed radio broadcasts.
  • Gold Standard: The authors created a gold standard for evaluating paraphrasing quality. This involved manually annotating metaphorical expressions in the BNC subset and then having human annotators evaluate the system's generated paraphrases.

Key Insights and Ideas

  • Reframing Metaphor Interpretation as Paraphrasing: This is the central innovation. By treating metaphor interpretation as generating literal replacements, the system bypasses the need for explicit metaphor knowledge bases and produces output directly usable by other NLP systems.
  • Leveraging Selectional Preferences for Literalness: The use of automatically induced selectional preferences is crucial for distinguishing between literal and metaphorical language, and for filtering out paraphrases that might still be figurative.
  • WordNet for Semantic Similarity: Utilizing WordNet's hierarchical structure to find shared meanings between metaphorical and literal terms is an effective way to constrain the search space and ensure semantic coherence.
  • High Accuracy without Hand-Crafted Knowledge: The system's ability to achieve high accuracy without relying on extensive, manually curated knowledge is a significant step towards more scalable and generalizable metaphor processing.

Key Results

  • The system achieved an accuracy of 0.81 in identifying correct literal paraphrases for metaphorical expressions in the evaluation setting 1.
  • In evaluation setting 2, the system's Mean Reciprocal Rank (MRR) was 0.63, outperforming a baseline MRR of 0.55.
  • The selectional preference-based reranking significantly improved accuracy by 26% over the baseline.

Limitations and Future Work

  • Limited Scope: The current system primarily focuses on single-word metaphors, specifically verb-subject and verb-object constructions.
  • Conventionality Issues: Highly conventionalized metaphors, where the metaphorical origin is clear but the sense is lexicalized, can still pose challenges for agreement and accurate interpretation.
  • Potential for Opposite Meanings: In some cases, the system might produce paraphrases with opposite meanings (e.g., "tension mounted" -> "tension lifted"), which the authors attribute to feature similarity components in WordNet.
  • Future Work: The authors express confidence that the interpretation techniques can be extended to other parts of speech and a wider range of syntactic constructions. Expanding the system to handle more complex phrases is a key direction for future research.

Summary updated

BibTeX
@inproceedings{shutova2010automatic,
  author = {Ekaterina Shutova},
  title = {{Automatic Metaphor Interpretation as a Paraphrasing Task}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2010},
}
Figure