Explainable Metaphor Identification Inspired by Conceptual Metaphor Theory
Authors: Mengshi Ge, Rui Mao, Erik Cambria10681-10689
AAAI 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our experiments show the effectiveness of the proposed model in metaphor identification, and concept mapping tasks, respectively. |
| Researcher Affiliation | Academia | Mengshi Ge*, Rui Mao*, Erik Cambria Nanyang Technological University, Singapore mengshi001@e.ntu.edu.sg, rui.mao@ntu.edu.sg, cambria@ntu.edu.sg |
| Pseudocode | No | The paper does not contain any clearly labeled pseudocode or algorithm blocks. |
| Open Source Code | No | The paper does not contain any explicit statement about releasing source code or a link to a code repository for the described methodology. |
| Open Datasets | Yes | Datasets MOH Shutova, Kiela, and Maillard (2016) developed a verb-noun pair dataset, parsed from the collection of Mohammad, Shutova, and Turney (2016). The dependent relationships between verb-noun pairs are either verb-subject or verb-direct object. We conduct 10-fold cross-validation with the MOH dataset for benchmarking. TSV (Tsvetkov et al. 2014) is an adjective-noun pair dataset. We randomly sample 200 word pairs from the original training set as the development set. GUT (Gutierrez et al. 2016) is an adjective-noun pair dataset. |
| Dataset Splits | Yes | We conduct 10-fold cross-validation with the MOH dataset for benchmarking. and We randomly sample 200 word pairs from the original training set as the development set. and The reported testing results and 10-fold cross-validation results are based on the model that achieves the highest F1 score on the development sets. |
| Hardware Specification | Yes | apart from MOH, 256 batch size runs out of memory on the other datasets, based on our employed model and Ge Force GTX 1080 Ti GPU. |
| Software Dependencies | Yes | The model depends on Cuda 9.2 (NVIDIA, Vingelmann, and Fitzek 2020), Pytoch 1.7.1 (Paszke et al. 2019), and optimized with Adam optimizer (Kingma and Ba 2014) and a learning rate of 1e-5. |
| Experiment Setup | Yes | The batch size for training MOH is 256, while the batch size for training TSV, GUT, and the combination of MOH and TSV is 1284. We train the model with 40 epochs5. The reported testing results and 10-fold cross-validation results are based on the model that achieves the highest F1 score on the development sets. The model depends on Cuda 9.2 (NVIDIA, Vingelmann, and Fitzek 2020), Pytoch 1.7.1 (Paszke et al. 2019), and optimized with Adam optimizer (Kingma and Ba 2014) and a learning rate of 1e-5. We use Ro BERTa-large as the encoder with a dropout rate of 0.3. In the common association acquisition procedure, we employ the 3 most frequent associations of a verb or an adjective in the Wikipedia dump. ϕ and γ in E.q. 7 are 0.05 and 0.005 respectively for balancing the losses between subtasks. |