--- license: apache-2.0 task_categories: - text2text-generation language: - bg pretty_name: Bulgarian Spelling Mistakes size_categories: - 10K - [Licensing Information](#licensing-information) ## Dataset Description - **Paper :** https://aclanthology.org/2024.icnlsp-1.38/ - **Point of Contact:** Bozhidar Klouchek; Please contact klouchek.bozhidar@gmail.com ### Dataset Summary This is a dataset of sentences in Bulgarian with spelling mistakes created by automatically inducing errors in correct sentences. ### Supported Tasks - `text2text-generation`: The dataset can be used to train a model for spelling error correction, which consists in correction of spelling errors. in a source sentence, resulting in a correct version. ### Languages - `bg`: Only Bulgarian is supported by this dataset. ## Dataset Structure ### Data Instances Each instance contains an `error_type`, which can be one of four pre-defined classes. This error_type describes the error found in the sequence in `erroeneous` which has been corrected in the sequence under `correct`. ``` { 'error_type': 'article_misuse', 'erroeneous': 'Възстанието влияе на всички ни!', 'correct': 'Въстанието влияе на всички ни!' } ``` ### Data Fields - `error_type`: a string sequence that can be one of: - **vowel_change** - **double_consonant** - **end_of_lemma_consonant** - **double_t_or_n** - **loss_of_t_or_d_sound** - **random_char** - **semantic** - `erroeneous`: a string sequence of the erroneous sentence containing up to three errors in the error-correct pair - `correct`: a string sequence of the correct sentence in the error-correct pair ### Data Splits No pre-defined split has been applied to the dataset so the developer has the freedom to choose one that suits the task. ## Dataset Creation ### Curation Rationale Main motivations for the creating of this dataset: - resources for error correction NLP systems in Bulgarian are scarce, so this dataset aims to enocourage the development and evaluation of more such systems, - solutions within Bulgarian NLP are traditional machine learning methods, so a dataset like this aims to encourage the development of state-of-the-art models (e.g. deep learning approaches) ### Source Data The source data for this dataset has been collected from Bulgarian Wikipedia articles. ### Initial Data Collection and Normalization The data collection process was such: 1) Data collection: Articles from Bulgarian Wikipedia were collected using Wikipedia's API. 2) The source texts underwent POS tagging and sentence segmentation using the tool from the work of [(Berbatova M., Ivanov F.,2023)](https://doi.org/10.60063/gsu.fmi.110.37-50). 3) Only sentences that (1) have three or more words, and (2) contain a token tagged as a verb were kept, as many of the sentences were simply article titles, links and other textual data we weren't interested in. ### Introducing Spelling Errors The dataset has been created by introduced spelling errors in correct sentences. Errors of the pre-defined types were induced using Python scripts, taking into account the error's nature, the algorithm is outlined as such: - Take a source sentence, which would be kept as the reference. - Induce an error of the pre-defined types, if that is possible (some sentences do not contain necessary pre-requisites for certain error types e.g. a relative pronoun is necessary to introduce an error of that type) - Pair up the correct and erroneous versions These errors have only been applied to lemmas: - larger than three characters, because shorter words tend to be functional words, and - not containg capitalised letters, on the assumption that it might be a named entity The resulting erroneous sentences were created with four different types of changes, reflected in the error_type column. In the examples below the first sequence is the **CORRECT** form, and the resulting sequence is **ERRONEOUS**. 1) **vowel_change**: a vowel that is not under stress is changed to its twin sounding vowel (e.g. 'кръг**ъ**л' -> 'кръг**а**л') 2) **double_consonant**: when two neighbouring consonants with differing sound strengths change the first one to match the second ('постано**в**ка' -> 'постано**ф**ка') 3) **end_of_lemma_consonant**: changing a strong sounding consonant to its weak sounding twin if at the end of a lemma ('маси**в**' -> 'маси**ф**') 4) **double_t_or_n**: removing one of the 'н' or 'т' letters when two of the same kind are found neighbouring, often in suffixes ('пролет**т**а' -> пролета, един**н**о -> едино) 5) **loss_of_t_or_d_sound**: removing either 'т' or 'д' if found in a lemma where the sound of the removed letter is lost ('вес**т**ник' -> 'весник', 'звез**д**на' -> звезна) 6) **random_char**: replacing one random character in a lemma with another random character ('мо**м**иче' -> 'мо**г**иче') 7) **semantic**: replacing, removing, adding a character or swapping two characters of a lemma such that the resulting word is a valid word syntactically, but the sentence is no longer semantically coherent ('Момчето, което обичам.' -> 'Момчето, ко**т**ето обичам.') ### Personal and Sensitive Information The source of this dataset is open-source data collections (Wikipedia) and carry the same amount of risk of personal and/or sensitive information as they do. ## Considerations for Using the Data ### Social Impact of Dataset A dataset like this can be beneficial for language learners and developers in the error correction community. ### Discussion of Biases The `error_type` classes are not distributed equally, as some errors are more common than others. It's important the developer utlising this dataset is aware of this as to not create error correction system/evaluations that are biased. ### Other Known Limitations There are many spelling errors not covered by this dataset, as it's the first of its kind. Hopefully, it encourages people to create more datasets like this and models that utilise them. ## Additional Information ### Licensing Information The license of the dataset is [apache2.0](https://www.apache.org/licenses/LICENSE-2.0).