The primary data of the present dataset originate from the following dataset, which functioned as the primary source of our Dataset.
- Source Intent Classification Dataset:
- Original Dataset: CLINC150
- Original Paper : An Evaluation Dataset for Intent Classification and Out-of-Scope Prediction
- License: Creative Commons Attribution 3.0 Unported License
For the development of our dataset, we selected the CLINC150 dataset because it is an IC dataset, offering a wide range of single intent utterances of different domains and intents that closely resemble human-system interactions. We utilize the full version of CLINC150 dataset and focus solely on in-scope examples, excluding the out-of-scope samples. Therefore, our dataset has been created based on the subset of the full dataset consisting of 22,500 in-scope utterances.
Our dataset, an expansion of the original CLINC150 dataset, incorporates multi-intent utterances containing the linguistic phenomena of Anaphora, Cataphora and Ellipsis. It is designed to handle multi-intent scenarios, encompassing double-intent and single-intent utterances. Thus, the dataset includes double-intent utterances with Anaphora, Cataphora & Ellipsis, double-intent utterances without these linguistic phenomena using structures as “Not only- but also, besides, furthermore, moreover” to produce more natural utterances, and single intent utterances. Furthermore, as the primary data were originated from CLINC150, the original utterances dataset were enriched with the aforementioned linguistic phenomena, occurring in double-intent utterances manually; we prefer to call them utterances, rather than sentences, because we view them as acts of speech. However, the challenges encountered in incorporating thephenomena of cataphora, anaphora, and ellipsis necessitated the modification of some of the initial utterances of the CLINC150 dataset, rather than adopted the straightforward conjunction of two utterances, we adhered to the rules governing anaphora, cataphora and ellipsis.
The file all_data.json is the full version of our Dataset. This file includes training, testing and validation sets and comprises 1,875 utterances, of which 910 exemplify the linguistic phenomena of anaphora, cataphora and ellipsis. The training set comprises a total of 1,315 utterances, of which 975 represent double intents, with 650 instances exhibiting linguistic phenomena and 325 instances without these linguistic phenomena. Additionally, the training set includes 340 single intent utterances. The test set consists of 280 utterances, with 195 representing double intents, comprising 130 examples with linguistic phenomena and 65 utterances without, along with 85 single-intent utterances. The validation set mirrors the test set in terms of composition, with 280 utterances, 195 double intent utterances (130 with linguistic phenomena and 65 without), and 85 single intent utterances.
See domains.json to add mapping of domains to intents.