rizki-syazali/itqa-flatten-header

Dataset Card for IndoHiTab (Flattened Version)

Dataset Description

IndoHiTab is a newly constructed, high-quality dataset specifically designed to solve the Table Question Answering (TQA) task for the Indonesian language. Due to the lack of publicly available resources in this domain, this dataset serves as the primary benchmark for evaluating extractive table parsers, such as the IndoTaPas model.

This specific dataset repository contains the Flattened Version of the primary, manually annotated IndoHiTab dataset. It was developed by translating a subset of the English HiTab dataset using human annotators under a standardized protocol. This is the exact preprocessed format utilized in our study to train and evaluate the TaPas architecture.

Dataset Structure

Dataset Statistics

After filtering out questions requiring header selection (due to TaPas architectural constraints), this manually annotated dataset contains a total of 2,559 instances, split into training and testing sets:

Dataset	Train	Test	Translation Method	Purpose
IndoHiTab	2,057	502	Manual (Human)	Main training and evaluation

Data Instances (Flattened Structure)

In this Flattened Format, tables with multi-level (hierarchical) column and row headers have been preprocessed and converted into a flat, 1-dimensional structure. This is achieved by concatenating the top-level header with lower-level headers. The resulting table includes a dedicated header array, and all string values (including the data matrix) are lowercased.

Example Instance:

{
  "id": "00d47c9aac5050539645dcae34f78570",
  "question_type": ["none"],
  "question": "berapa persen penduduk canada dengan pendapatan rumah tangga sebesar $150,000 atau lebih yang mengatakan bahwa pengaruh hoki terhadap identitas nasional sangat penting?",
  "answer_coordinates": [[21, 5]],
  "answer_text": ["49"],
  "table_source": "statcan",
  "table": {
    "data": [
      ["total", "70", "69", "64", "55", "46"],
      ["kelompok umur", "", "", "", "", ""],
      ["kelompok umur 15-24", "70", "56", "51", "48", "47"],
      ["kelompok umur 25-34", "71", "65", "59", "51", "50"],
      ["tingkat pendidikan tertinggi yang ditamatkan di bawah sekolah menengah", "65", "68", "66", "60", "46"],
      ["..."]
    ],
    "header": [
      "karakteristik sosio-demografi dan ekonomi",
      "piagam hak dan kebebasan persen",
      "bendera persen",
      "lagu kebangsaan persen",
      "rcmp persen",
      "hoki persen"
    ],
    "id": "2613",
    "title": "persepsi simbol nasional sebagai hal yang sangat penting untuk identitas kanada, berdasarkan karakteristik sosio-demografi dan ekonomi, 2013"
  }
}

Data Fields

• id: Unique string identifier for the question-table pair.
• question_type: List of strings indicating the type of question.
• question: The natural language question in Indonesian.
• answer_coordinates: List of lists containing the [row_index, column_index] of the exact answer(s) within the flattened data matrix.
• answer_text: List of strings containing the expected answer value(s).
• table_source: Source of the original table (e.g., statcan, totto, nsf).
• table: A dictionary representing the flattened table containing:
  • header: A 1-dimensional array of column headers.
  • data: A 2-dimensional array (matrix) of the table's row contents.
  • id: The table identifier.
  • title: The title or caption of the table.

How to Use

You can easily load this dataset using the datasets library from Hugging Face:

from datasets import load_dataset

# Load the dataset
dataset = load_dataset("rizki-syazali/itqa-flatten-header")

# Print the first instance of the training set
print(dataset['train'][0])