This dataset is a parallel-script dataset for Sindhi, containing the same Sindhi sentences represented in multiple writing systems and transliteration schemes.

The dataset contains four corresponding columns:

devanagari: Sindhi sentences written in Devanagari script. This serves as the ground-truth source representation.
persoarabic: Sindhi sentences represented in the Perso-Arabic script.
khudabadi: Sindhi sentences represented in the Khudabadi script.
roman: Sindhi sentences written using Roman script.

The dataset is intended for research and development involving Sindhi script conversion, transliteration, Romanization, normalization, and multiscript NLP.

Dataset Structure

Column	Description
`devanagari`	Ground-truth Sindhi sentence written in Devanagari
`persoarabic`	Sindhi sentence represented in Perso-Arabic script
`khudabadi`	Sindhi sentence mapped from Devanagari to Khudabadi
`roman`	Sindhi sentence represented in Roman script based on real-world writing patterns

Each row represents the same underlying Sindhi sentence across all four representations.

Data Creation

Devanagari

The devanagari column contains Sindhi sentences written in Devanagari script and serves as the ground-truth representation for the dataset.

Devanagari was selected as the ground-truth script because Sindhi written in Devanagari explicitly represents both consonants and vowels, providing information that is not always explicitly represented in the Perso-Arabic script. In addition, the Devanagari text follows standardized Sindhi spelling conventions.

This makes Devanagari a reliable and consistent source for deterministic script conversion. Its standardized spelling also makes rule-based manipulation considerably less ambiguous than Roman Sindhi, where the same Sindhi word can be written in many different ways by different users.

The other representations in the dataset were generated from the Devanagari sentences.

Perso-Arabic

The persoarabic column was initially generated using a rule-based transliteration system.

The resulting word spellings were then post-corrected using:

A Sindhi dictionary to generate and identify appropriate spellings for individual words.
A language model to select or improve word spellings based on the surrounding linguistic context.

The correction process operates at the word level rather than rewriting the sentence as a whole, preserving the original sentence structure and content.

This additional correction was necessary because Sindhi Perso-Arabic has multiple letters that can represent the same or very similar sounds. For example, the sound represented by s can be written using س, ص, or ث, while a z sound can be represented using ز, ذ, ض, or ظ. Therefore, a direct character-level transliteration from Devanagari cannot always determine the appropriate Perso-Arabic spelling for a word.

The dictionary and language model are therefore used to determine the appropriate word-level spelling based on the word and its surrounding context, rather than simply applying a character-to-character mapping.

The resulting Perso-Arabic text combines rule-based transliteration with dictionary-based and language-model-based word-level correction.

Khudabadi

The khudabadi column is generated through a one-to-one character mapping from Devanagari to Khudabadi.

Each Devanagari character is mapped to its corresponding Khudabadi character according to a predefined character mapping.

Unlike the Perso-Arabic and Roman representations, this process does not involve language-model-based correction or modeling of writing behavior. The mapping is deterministic.

Roman

The roman column represents Sindhi written in Roman script and is generated from the Devanagari representation.

Devanagari is used as the source because it explicitly represents both consonants and vowels, providing the linguistic information required to generate the Roman representation.

The Romanization itself is rule-based, but the rules were designed after modeling patterns observed in real user-written Roman Sindhi. Rather than enforcing a single standardized Romanization system, the rules attempt to reproduce the different ways Sindhi speakers actually write Sindhi using Roman characters.

For example, the same Sindhi sound or word may be represented in multiple ways by different users. The Romanization system therefore models this variation and allows the resulting Roman text to reflect diverse real-world writing patterns.

Thus, Devanagari provides the underlying linguistic information, while real-world Roman Sindhi writing behavior is used to determine how that information is represented in Roman script.

Why Devanagari Is the Ground Truth

The choice of Devanagari as the ground-truth representation is based on its suitability as a consistent source representation for Sindhi.

Sindhi Devanagari provides explicit representation of both consonants and vowels, while Perso-Arabic Sindhi can have multiple possible letters for the same sound. In addition, the dataset uses standardized Devanagari spellings.

This provides a more deterministic foundation for generating the other representations.

Roman Sindhi is also generated from Devanagari because its explicit vowel and consonant information provides a consistent source for Romanization. However, the Roman output intentionally models the variation found in real-world Roman Sindhi rather than enforcing a standardized spelling system.

Therefore, the Devanagari column should be considered the reference representation, while the other columns are derived representations produced using different conversion methodologies.

Intended Uses

The dataset can be used for:

Sindhi script conversion
Sindhi transliteration
Devanagari to Perso-Arabic conversion
Perso-Arabic to Devanagari conversion
Devanagari to Khudabadi conversion
Sindhi Romanization
Multiscript Sindhi NLP
Sequence-to-sequence model training
Transliteration model evaluation
Sindhi text normalization
Studying variation in Roman Sindhi
Script identification and conversion

Data Processing Summary

The overall data generation pipeline can be summarized as:

                 ┌─────────────────────┐
                 │  Sindhi Devanagari  │
                 │    Ground Truth     │
                 └──────────┬──────────┘
                            │
              ┌─────────────┼─────────────┐
              │             │             │
              ▼             ▼             ▼
       Perso-Arabic     Khudabadi       Roman
              │             │             │
              ▼             │             ▼
       Rule-based           │      Rule-based
       transliteration      │      transliteration
              │             │             │
              ▼             │             ▼
       Word-level           │      Rules based on
       dictionary +         │      real-user Roman
       language-model       │      writing patterns
       correction           │             │
              │             │             │
              └─────────────┴─────────────┘

Limitations

The persoarabic, khudabadi, and roman columns are derived representations rather than independently collected human transcriptions.

The persoarabic column has undergone rule-based generation followed by dictionary and language-model-based word-level correction, so errors or ambiguities may remain.

The khudabadi column is based on a deterministic character mapping and therefore represents the defined mapping rather than necessarily capturing every historical or orthographic variation of Khudabadi writing.

The roman column intentionally models variation found in real-world Roman Sindhi. Consequently, it should not be interpreted as a standardized Romanization system. Multiple valid Roman representations can exist for the same Sindhi sentence.

Quality Considerations

The Devanagari representation is treated as the reference representation because of its explicit vowel and consonant representation and standardized spelling.

The derived representations should therefore be evaluated against the Devanagari source while also accounting for the characteristics of each writing system. In particular, differences in Roman Sindhi should not automatically be considered errors when they correspond to common user writing patterns.

License

This dataset is licensed under the Creative Commons Attribution 4.0 International (CC BY 4.0) license.

You are free to share, copy, modify, and use the dataset, including for commercial purposes, provided that appropriate credit is given to:

Fahad Maqsood Qazi

Attribution should include the dataset name, creator name, and a link to the original dataset repository where reasonably possible. If modifications are made, they should be indicated.

CC BY 4.0 License

Downloads last month: 37

Total file size:

148 MB