Weeding out Conventionalized Metaphors: A Corpus of Novel Metaphor Annotations

Abstract

We encounter metaphors every day, but only a few jump out on us and make us stumble. However, little effort has been devoted to investigating more novel metaphors in comparison to general metaphor detection efforts. We attribute this gap primarily to the lack of larger datasets that distinguish between conventionalized, i.e., very common, and novel metaphors. The goal of this paper is to alleviate this situation by introducing a crowdsourced novel metaphor annotation layer for an existing metaphor corpus. Further, we analyze our corpus and investigate correlations between novelty and features that are typically used in metaphor detection, such as concreteness ratings and more semantic features like the Potential for Metaphoricity. Finally, we present a baseline approach to assess novelty in metaphors based on our annotations.

Predicting Humorousness and Metaphor Novelty with Gaussian Process Preference Learning

“…We repeat Task 2 with different subsets of these features to determine the most effective combination. The token frequency feature has previously been shown to distinguish between metaphoric and literal use (Beigman Klebanov et al, 2014), but also to be indicative of metaphor novelty (Do Dinh et al, 2018). By incorporating the polysemy feature we seek to increase performance especially for the funniness dataset, which includes many puns.…”

“…Metaphor Novelty Dataset. We use the metaphor novelty dataset of Do Dinh et al (2018), which contains novelty scores for metaphors (i.e., metaphoric tokens) from the VU Amsterdam Metaphor Corpus (Steen et al, 2010) across four genres: news, fiction, conversation transcripts, and academic texts. The metaphors were compared by crowd workers using best-worst scaling tuples of four randomly chosen metaphors -that is to say, annotators were presented with random selections of four sentences with the metaphoric tokens highlighted, and they selected the most novel and most conventionalised metaphors from this set.…”

“…Because not all pairs are unique, and different pairs were extracted for different annotators, the number of unique pairs decreases, and the number of annotations per unique pair is less than three. We also use the gold standard provided by Do Dinh et al (2018), which was obtained by applying BWS to the complete dataset. Table 4 presents some statistics on the humour and metaphor novelty datasets.…”

“…girls often produce responses like 'often go through a bad patch for a year' 'when you tried to read the book, there was nothing there, because the words started as a coat-hanger to hang pictures on.' Table 2: Examples of statements from the Metaphor Novelty dataset (Do Dinh et al, 2018) containing highlighted metaphors. The upper example is highly conventionalised, while the lower is more novel and creative.…”

Filter and Annotate: Towards Automatic Identification of Genuine Metaphoricity

“…In contrast, there is no such dataset for genuine metaphoricity at all. Closest to our use case are the English datasets annotated for novel metaphor relations [10] and novel metaphor tokens [11].…”