Update README.md

README.md

@@ -26,13 +26,9 @@ NewsMTSC is a high-quality dataset consisting of more than 11k manually labeled
 ## Dataset
 
 ### Files
-The dataset consists of three splits. We suggest to use the splits as follows:
+The dataset consists of two subsets (`rw` and `mt`), each consisting of three splits (train, validation, and test). We recommend to use the `rw` subset, which is also the default subset. Both subsets share the same train set, in which the three sentiment classes have similar frequency since we applied class boosting. The two subsets differ in their validation and test sets: `rw` contains validation and test sets that resemble real-world distribution of sentiment in news articles. In contrast, `mt`'s validation and test sets contain only sentences that each have two or more (different) targets, where each target's sentiment was labeled individually. 
 
-* `train.jsonl` - For **training**.
-* `devtest_rw.jsonl` - To evaluate a model's classification performance on a "**real-world**" set of sentences, i.e., the set was created with the objective to resemble real-world distribution as to sentiment and other factors mentioned in the paper.
-* `devtest_mt.jsonl` - To evaluate a model's classification performance only on sentences that contain at least **two target mentions**. Note that the mentions were extracted to refer to different persons but in a few cases might indeed refer to the same person.
-
-**For most projects, we suggest** to use `train.jsonl` for training and `devtest_rw.jsonl` for evaluation. More information can be found in our [paper](https://aclanthology.org/2021.eacl-main.142.pdf).
+More information on the subsets can be found in our [paper](https://aclanthology.org/2021.eacl-main.142.pdf).
 
 ### Format
 Each split is stored in a JSONL file. In JSONL, each line represents one JSON object. In our dataset, each JSON object consists of the following attributes. When using the dataset, you most likely will need (only) the attributes highlighted in **bold**.