Update README.md

README.md

@@ -9,24 +9,64 @@ size_categories:
 # DICE: Dataset for Controlled Evaluation of Idiomatic Expressions
 
 ## Summary
-DICE is a corpus for testing models' understanding of potentially idiomatic expressions (PIEs), and potential effects of memorisation.
+DICE is a corpus for testing models' understanding of potentially idiomatic expressions (PIEs) in contexts, and potential effects of memorisation.
 
 PIEs are expressions that can be interpreted with a figurative meaning or a literal meaning, depending on the context in which they occur.
 For instance, "let the cat out of the bag" could mean "revealing a secret", it could also hold a more literal interpretation of "releasing a cat from a bag". 
 
-Improving on previous datasets for idiomaticity detection, we keep the form of the expression the same in both literal and figurative contexts. In doing so, we are able to test the robustness of the models. If model do understanding context, they would be able to correctly resolve the interpretation of the expressions.
+Improving on previous datasets for idiomaticity detection, we keep the form of the expression the same in both literal and figurative contexts. In doing so, we are able to test the robustness of the models in their understanding of context. If model do understanding context, they would be able to correctly resolve the interpretation of the expressions.
 
 ## Dataset Structure
 
 Per instance example:
 
+ID,Idiom,Sentence_Original,Sentence_Cleaned,Extracted_Phrase,label
+
+ID_i_187a,scratch the surface,"So far , research into psychoneuroimmunology has done no more than scratch the surface of this potentially important topic .","So far, research into psychoneuroimmunology has done no more than scratch the surface of this potentially important topic.",scratch the surface,i
+
 ```json
 {
-    'ID': '',
-    'idiom': 'across the board',
-    'sentence': '',
-    'cleaned_sentence'
-    'label': 'figurative',
+    'ID': 'ID_i_187a',
+    'Idiom': 'scratch the surface',
+    'Sentence_Original': 'So far , research into psychoneuroimmunology has done no more than scratch the surface of this potentially important topic .',
+    'Sentence_Cleaned': 'So far, research into psychoneuroimmunology has done no more than scratch the surface of this potentially important topic.',
+    'Extracted_Phrase': "scratch the surface",
+    'label': 'i',
 }
 ```
 
+The fields are the following:
+
+- `ID`: The ID of the sentence, where the middle letter represents the class (`i`: figurative, `l`:literal).
+- `Idiom`: The base expression.
+- `Sentence_Original`: The sentence containing the PIE. Note, if the sentence was taken from MAGPIE, there would be a single preceding whitespaces before quotes, fullstops and other punctuation marks.
+- `Sentence_Cleaned`: The sentence containing the PIE, where the whitespace and tokenisation has been cleaned up.
+- `Extracted_Phrase`: The exact form of the PIE, as it appears in the sentence. For instance, the base expression might be "busy bee", but the expression appears as "busy bees" in the sentence.
+- `label`: Gold label representing the correct interpretation (`i`: figurative, `l`:literal).
+
+
+### Data Splits
+
+|     Label| No. Sentences|
+|----------:|-----:|
+|`l`   | 1033 |
+|`i`   | 1033 |
+
+We also balance the expressions across each PIE. For instance, if there are two sentences for "spill the beans" in literal contexts, there are also an equal number of "spill the beans" in the figurative context.
+
+
+## Citation Information
+
+@misc{mi2025rollingdiceidiomaticityllms,
+      title={Rolling the DICE on Idiomaticity: How LLMs Fail to Grasp Context}, 
+      author={Maggie Mi and Aline Villavicencio and Nafise Sadat Moosavi},
+      year={2025},
+      eprint={2410.16069},
+      archivePrefix={arXiv},
+      primaryClass={cs.CL},
+      url={https://arxiv.org/abs/2410.16069}, 
+}
+
+
+
+