Deexaggeration
Authors: Li Kong, Chuanyi Li, Vincent Ng
IJCAI 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In this paper, we lay the groundwork for the computational study of understanding hyperbole by (1) defining a structured representation to encode what is being exaggerated in a hyperbole in a non-hyperbolic manner, (2) annotating the hyperbolic sentences in two existing datasets, HYPO and HYPO-cn, using this structured representation, (3) conducting an empirical analysis of our annotated corpora, and (4) presenting preliminary results on the deexaggeration task. |
| Researcher Affiliation | Academia | 1State Key Laboratory for Novel Software Technology, Nanjing University, Nanjing, China 2Human Language Technology Research Institute, University of Texas at Dallas, Richardson, Texas, USA |
| Pseudocode | No | The paper describes the approaches for static vs. dynamic prediction and tuple prediction, but it does not include any structured pseudocode or algorithm blocks. |
| Open Source Code | No | To stimulate further work on deexaggeration, we make our annotations publicly available.7The annotations can be downloaded from http://lichuanyi.info/files/papers/Deexaggeration.xlsx. This only refers to annotations, not the source code for the methodology. |
| Open Datasets | Yes | For annotation, we use two corpora previously assembled for research on automatic hyperbole detection, HYPO (English), and HYPO-cn (Chinese). To stimulate further work on deexaggeration, we make our annotations publicly available.7The annotations can be downloaded from http://lichuanyi.info/files/papers/Deexaggeration.xlsx. |
| Dataset Splits | Yes | All results are obtained via five-fold cross-validation experiments.6 In each fold experiment, we use three folds for training, one fold for development, and one fold for testing. |
| Hardware Specification | No | The paper does not provide specific details about the hardware used for running the experiments, such as GPU or CPU models. |
| Software Dependencies | No | The paper mentions using specific BERT models (albert tiny, bert small) and Microsoft's Uni LM, but it does not provide version numbers for these or any other software dependencies. |
| Experiment Setup | Yes | For all of the models described above, the maximum length of the input is set to 20 (measured in words for English and characters for Chinese). The batch size is set to 8. The number of epochs is searched out of {20, 25, 30, 35} using grid search on the development set. To train the sequence taggers and classifiers described above (i.e., the BERT-based static/dynamic classifier, the BERT-Bi LSTM-CRF model, and the BERT classifier for predicting an element s value), we use negative cross-entropy as the loss function. The initial learning rate is set to the default value of 0.001. Re LU is chosen as the activation function in the fully-connected layer. Two parameters are tuned using grid search on the development set: dropout (searched out of {0.1, 0.2, 0.3, 0.4}) and the optimizer (SGD or RMSprop). To train the seq2seq model, we use cross-entropy as the loss function and Adam as the optimizer with an initial learning rate of 0.001. |