Atis-Aligned is a semantic parsing dataset that was obtained by augmenting the popular ATIS dataset with word alignments.
The dataset can be found in the data folder and it comes in .csv format. There are five columns:
- ID : contains the id of the example
- NL : contains the natural language queries
- MR : contains the meaning representation
- ALIGNMENT : contains the word alignments
- MONOTONIC : contains 0-1 labels indicating whether the alignment is monotonic
There are two versions of the dataset:
EN.csvcontains the dataset with manually annotated word alignmentsEN_giza.csvcontains the dataset with automatically annotated word alignments using the GIZA++ tool
The splits folder contains lists of IDs for three different test splits:
- a question split, which ensures that the NL questions at test time have not been observed at training time
- a query split, which ensures that the MR queries at test time have not been observed at training time
- a length split, which ensures that the MR queries at test time are longer than those observed at training time
The test set IDs are found in the test.txt files. Moreover, we provide three development sets dev1.txt, dev2.txt and dev3.txt, which can be used for validation.