---
language: san
license: apache-2.0
tags:
- dependency-parsing
- nested-compound-type-identification
- lstm
- nlp
- sanskrit
- pytorch
---

# DepNeCTI-LSTM: Dependency-based Nested Compound Type Identification for Sanskrit

This repository contains the datasets proposed in the paper [Sandhan et al., 2023](https://arxiv.org/abs/2310.09501)

---
# Summary

The **NeCTIS** dataset is created for **nested compound type identification in Sanskrit**, focusing on multi-component compounds (especially with more than 2 components). It includes **coarse** and **fine-grained** semantic type annotations.

### Key Features

- **Two datasets**:
  - **NeCTIS**: In-domain (Prose)
  - **NeCTIS-OOD**: Out-of-domain (Poetry)

- **Annotations**:
  - **Coarse-level**: 4 broad compound types:
    - **Avyayībhava** (Indeclinable)
    - **Bahuvrīhi** (Exocentric)
    - **Tatpurusha** (Endocentric)
    - **Dvandva** (Copulative)
  - **Fine-grained**: 86 detailed sub-types

### Dataset Statistics

| Dataset        | #Nested Compounds | Train  | Test  | Dev   | Compound Types |
|----------------|-------------------|--------|-------|-------|----------------|
| **NeCTIS**     | 17,656            | 12,431 | 3,493 | 2,405 | 4 (86)         |
| **NeCTIS-OOD** | 1,189             | —      | 1,189 | —     | 4 (86)         |


### Domain & Genre

- **NeCTIS**: Philosophical, Literary, and Ayurvedic domains → *Prose*
- **NeCTIS-OOD**: Paurāṇic (epic literature) domain → *Poetry*

  - Poetry tends to include more novel, complex, and metrical compounds.

### Annotation Process

- Funded by **DeitY (2009–2012)** as part of the **Sanskrit-Hindi Machine Translation project**
- Annotation done by **6 institutes**, each with ~10 members across 3 expertise levels:
  - **Junior Linguist** (Master’s in Sanskrit)
  - **Senior Linguist** (Ph.D. in Sanskrit)
  - **Professional Linguist** (Professors)
- Multi-level quality checks and **cross-institute validation**
- Annotation guidelines based on **Pāṇinian grammar** and traditional commentaries.

**For Detailed information refer to the original paper.**


# Files
```
├── README.md
├── LICENSE
├── No Context CSV files/
│   ├── Combined.csv
│   ├── test.csv
│   ├── train.csv
│   ├── dev.csv
│   └── outofDomain.csv
├── With Context CSV files/
│   ├── Combined.csv
│   ├── test.csv
│   ├── train.csv
│   ├── dev.csv
│   └── outofDomain.csv

```

# Dive in the Dataset

## With Context 
- Total rows: **15,940 rows**.
- Split: Train (69%), Test (18%), Dev (13%).
- **train CSV**: **11,000 rows**.
- **test CSV**: **2,940 rows**.
- **dev CSV**: **2,000 rows**.
- The **combined CSV**: all above files merged, 15,940 rows total.
- These four CSV files include the following columns:  
  `Unnamed: 0`, `Raw_Tagged`, `Clean`, `Bio_tagged`, `Span_Tagged`, `Coarse_tag`, `Compound_lengths`, `Coarse_Span_Tagged`.
- The **out of Domain** contains **1139 rows**
- The **out-of-domain** dataset has an additional column, `Book`, and lacks the `Compound_lengths` column.


| Column Name        | Description                                                                                               |
|--------------------|-----------------------------------------------------------------------------------------------------------|
| **Unnamed: 0.1**   | Auto-generated row index by pandas when reading CSV, used as a unique identifier for each row.            |
| **Unnamed: 0**     | Another index column created during CSV operations, often redundant with `Unnamed: 0.1`.                   |
| **Raw_Tagged**     | Original input text with embedded annotation tags marking nested compound components and their types. Tags use angled brackets `< >` and suffix codes to show component boundaries and types. |
| **Clean**          | Cleaned, tokenized version of the text without any annotation tags or special characters, for plain reading. |
| **Bio_tagged**     | BIO tagging scheme at the token level indicating compound boundaries:   `B-C` = Beginning of a compound segment, `I-C` = Inside a compound segment, `O` = Outside any compound (non-compound token)                                                         |
| **Span_Tagged**    | Token spans marking compound segments along with their labels. Format: `start,end Label`. Multiple spans are separated by `\|`. For example, `0,2 BvS\|2,5 Ds` means tokens from index 0 to 2 form a `BvS` compound, tokens 2 to 5 form a `Ds` compound. |
| **Coarse_tag**     | Coarse-grained compound type annotations embedded with original tagged segments, indicating linguistic compound types such as `Bahuvrihi`, `Tatpurusha`, or `Dvandva`.  |
| **Compound_lengths** | List indicating the length (number of tokens) of each compound segment in the text. For example, `[1, 1]` means there are two compounds each one token long.   |
| **Coarse_Span_Tagged** | Combines span indices with coarse compound type labels to specify the token ranges and their compound categories. For example: `0,2 Bahuvrihi|2,5 Dvandva`. |
|**Book**| Present in ood dataset. Informs about the book from which the sentence is taken.|


## Without Context
- Total rows: **15,940 rows**.
- Split: Train (69%), Test (18%), Dev (13%).
- **train CSV**: **11,000 rows**.
- **test CSV**: **2,940 rows**.
- **dev CSV**: **2,000 rows**.
- The **combined CSV**: all above files merged, 15,940 rows total.
- The **out of Domain** contains **1139 rows**
- These five CSV files include the following columns:  
  `Unnamed: 0`, `Raw_Tagged`, `Clean`, `Bio_tagged`, `Span_Tagged`, `Coarse_tag`, `Coarse_Span_Tagged`.

| Column Name        | Description                                                                                               |
|--------------------|-----------------------------------------------------------------------------------------------------------|
| **Unnamed: 0.1**   | Auto-generated index column, unique identifier for each row.                                              |
| **Unnamed: 0**     | Another index column created during CSV processing, often redundant with `Unnamed: 0.1`.                   |
| **Raw_Tagged**     | Original text with embedded annotation tags marking nested compound components. Tags use angled brackets `< >` with suffix codes for component boundaries and types. Example: `<sa-sarzapaM>BvS <tumburu-DAnya-vanyaM>Ds`  |
| **Clean**          | Cleaned and tokenized text with no annotation tags, representing plain text tokens.                        |
| **Bio_tagged**     | BIO scheme token-level tagging indicating compound boundaries: `B-C` = Beginning of a compound segment `I-C` = Inside a compound segment `O` = Outside any compound segment.                                                                   |
| **Span_Tagged**    | Token span annotations marking compound segments with labels. Format: `start,end Label`. Multiple spans are separated by `\|`. For example, `0,2 BvS\|2,5 Ds` indicates tokens 0 to 2 form a `BvS` compound, and tokens 2 to 5 form a `Ds` compound.  |
| **Coarse_tag**     | Coarse-grained compound type annotations embedded in the original tagged text, indicating linguistic compound types such as `Bahuvrihi`, `Tatpurusha`, or `Dvandva`.      |
| **Coarse_Span_Tagged** | Combines token span indices with coarse compound type labels, e.g., `0,2 Bahuvrihi\|2,5 Dvandva` specifies token ranges and their corresponding compound categories.    |


# Paper
```
@misc{sandhan2023depnecti,
      title={DepNeCTI: Dependency-based Nested Compound Type Identification for Sanskrit}, 
      author={Jivnesh Sandhan and Yaswanth Narsupalli and Sreevatsa Muppirala and Sriram Krishnan and Pavankumar Satuluri and Amba Kulkarni and Pawan Goyal},
      year={2023},
      eprint={2310.09501},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}
```
Original paper [DepNeCTI: Dependency-based Nested Compound Type Identification for Sanskrit](http://arxiv.org/abs/2310.09501)

Github Repository of [DepNeCTI](https://github.com/yaswanth-iitkgp/DepNeCTI)