GlotWeb

About GlotWeb

GlotWeb is an advanced web indexing system specifically designed to address the digital resource gap for minority languages. Our system:

• Identifies web content in 402+ languages through multi-source aggregation
• Validates linguistic accuracy using GlotLID language identification
• Filters content to ensure quality while minimizing religious bias
• Compiles 169,155+ verified web links (47% in languages absent from major datasets)

Key Features

✔ Covers languages missing from FLORES-200, MADLAD-400, and Glot500
✔ Open-source pipeline with reproducible results
✔ Interactive demo showcasing language resources (Click Image below to access)

Getting Started

This documentation walks through GlotWeb's 4-step pipeline:

1. Search Setup: Configure and run web searches
2. Seed Generation: Filter initial results
3. Crawling: Expand and validate links
4. Cleaning: Deduplicate and finalize outputs

How to Use This Guide

1. Follow steps sequentially (1 → 4)
2. Each section includes:
   • Purpose explanation
   • Configuration options
   • Execution commands
   • Expected outputs
3. Requires basic Python/Docker knowledge

Tip: For quick setup, clone the repository and use the provided configuration files as templates.

Ready to begin? Proceed to Step 1: Set up SearXNG and perform search.

Step 1: Set up SearXNG and perform search

SearXNG

Install

export PORT=8080

docker pull searxng/searxng
docker run --rm \
             -d -p ${PORT}:8080 \
             -v "${PWD}/searxng:/etc/searxng" \
             -e "BASE_URL=http://localhost:${PORT}/" \
             -e "INSTANCE_NAME=my-instance" \
             searxng/searxng

add JSON output format:

In the ${PWD}/searxng directory, there is a file named settings.yml. In that file, you need to enable the JSON output format in the SearX configuration under the search.formats key like this:

search:
  formats:
    - html
    - json

modify uwsgi.ini : In the ${PWD}/searxng directory, there is a file named uwsgi.ini. In that file you need to modify buffer-size. Default is 8k. Increasing to 9k sometimes help with 'Internal Error 500'.

Default Value:

buffer-size = 8192

Change to:

buffer-size = 9216

Search Service Script: search_service.py

This is an object-oriented Python script that leverages the Searx API to perform searches and save the results to JSON files. The script is configurable using a YAML configuration file called 'search_config.yaml'.

Features

• Uses SearxSearchWrapper for querying multiple search engines.
• Handles retries for failed requests.
• Configurable search parameters through a YAML file.
• Configurable input file, search range, output directory, and other parameters.
• Automatically saves results in a structured JSON format.

Configuration file parameters:

searx_host: "http://127.0.0.1:8080"  # Searx instance URL
engines:
  - "bing"
  - "yahoo"
  - "qwant"
  - "duckduckgo"  # Search engines to be used
num_results: 50  # Number of results to fetch for each query
max_retries: 3  # Maximum number of retries for failed requests
retry_wait_time: 2  # Wait time (in seconds) between retries
output_file_prefix: "results"  # Prefix for output file names
output_directory: "search_dump"  # Directory to save output files
input_file: "input.txt"  # Path to the input file containing search queries
start_index: 0  # Start index for queries to process
end_index: 10  # End index for queries to process

Input file format:

The input file should be a tab-separated file where each line contains an ISO code and a sentence for search:

ISO_CODE_1    Search query 1
ISO_CODE_2    Search query 2

aa	Itiyobbiyah agattinoona sittal xayyossa yangalen qaadoodih baari gablusaanamah angicille le.
aai	Baise orot ta’ita’imon matah toniwa’an bar hinanutitiy gewas hinawowab.
aak	O xewanɨŋo na'nɨ re rɨnɨŋɨnigɨnɨ, ‘A'mɨna' sea'yɨ e imo'nɨŋa' wonɨrɨnɨ.’

The iso code for the input text file can be either 2 lettered format or 3 lettered format.

Usage

Run the script using: pwd should be be root of the directory.

python pipeline/search_service.py

The search results will be saved in the specified output directory (e.g., search_dump) as JSON files named according to the specified prefix and index range, e.g., results_0-10.json.

Customization

You can easily adjust the following parameters in the config.yaml file:

• Search engines: Add or remove engines in the engines list.
• Search range: Modify start_index and end_index to control which lines in the input file are processed.
• Output directory: Change output_directory to save results in a different location.

Step 2: Filter and generate seeds

Overview

This script filters web search dump/results based on domain restrictions, scrapes web pages, and performs language identification a FastText model for which we chose GlotLID. The processed data is stored in JSON format categorized by predicted languages.

Prerequisites

Dependencies

Ensure you have the following Python packages installed:

pip install fasttext trafilatura urllib3 tqdm pyyaml

Configuration already provided in the repository and must be changed according to user preferneces. Examples below:

model_path: "path/to/fasttext/model"
domain_file: "path/to/domain_filter.txt"
json_filename: "path/to/input.json"
iso_list_file: "path/to/iso_list.json"
output_directory: "path/to/output"

Running the Script

Execute the script with: pwd should be be root of the directory.

python pipeline/language_filter.py

Step 3: Search and scrape with seeds

Overview

This step takes the filtered seed URLs from Step 2 and performs deep crawling to discover additional web pages in the target languages. It includes:

• Web crawling from seed URLs
• Language detection using FastText (GlotLID)
• Domain filtering
• Parallel processing for efficiency
• Comprehensive logging and metadata collection

Prerequisites

Dependencies

Ensure you have the following Python packages installed:

pip install fasttext beautifulsoup4 requests trafilatura tqdm pyyaml urllib3

Configuration

The script uses config.yaml with these key parameters:

seed_crawler:
  max_pages: 100            # Maximum pages to crawl per language
  max_time: 3600              # Maximum crawling time in seconds
  crawl_delay: 1              # Delay between requests
  to_visit_growth_factor: 50   # Threshold for detecting circular links
  max_workers: 4              # Threads for parallel processing

url_settings:
  request_timeout: 10         # Timeout for web requests
  max_url_length: 65000         # Maximum URL length to consider

language_detector:
  model_path: "path/to/model" # Path to FastText model
  minimum_confidence: 0.7     # Minimum language confidence score
  desired_language: "bpy_Beng"      # Target language code
  save_text: False            # Whether to save scraped text

output:
  directory: "output"         # Output directory
  output_file_name: "{language}_filtered.json"  # Output filename pattern

batch_processing:
  enabled: False              # Enable batch mode
  input_labels: []            # List of language codes for batch
  cooldown_between_languages: 60  # Cool-down between languages

Input Requirements

Input JSON files from Step 2 (named as [LANGUAGECODE_SCRIPT].json)

Each JSON file should contain entries with:

• link: URL string

• lid_confidence: Confidence score (float)

• predicted_lid: Language code

Output

For each processed language, the script generates:

[LANGUAGE_CODE]_filtered.json - Filtered URLs with metadata

meta_data/[LANGUAGE_CODE]_meta_data.json - Crawling statistics including:

Seed URLs used

All discovered links

Filtered links

Unique new links

Rejected links

Usage

Single Language Processing

python pipeline/seed_crawler.py

Configure desired_language in config.yaml first.

Batch Processing:

Enable batch mode in config.yaml:

batch_processing:
  enabled: True
  input_labels: ["syl_Sylo", "bpy_Beng", "akh_Latn"]  # Your target languages

Run the same command:

python pipeline/seed_crawler_beta.py

Customization Options

• Crawling Behavior: Adjust max_pages and max_time to control crawling scope Modify crawl_delay to be more/less aggressive
• Language Detection: Change minimum_confidence for stricter/looser filtering
• Set save_text: True to store scraped content
• Performance: Increase max_workers for faster processing (requires more CPU) Adjust cooldown_between_languages for batch processing

Output:

• Change output directory and filename patterns
• Metadata collection is always enabled

Notes

• The script automatically skips domains listed in your domain filter file
• Progress bars are enabled by default (can be disabled in config)
• Comprehensive logging helps troubleshoot issues

Step 4: Filtering and Deduplication

Step 4.1: Domain Filtering

Purpose

This script performs final domain filtering on crawled results to exclude unwanted domains from both the main output and metadata files.

Key Features

• Loads crawled data and metadata JSON files
• Applies domain filtering using the configured domain blocklist
• Updates all metadata statistics after filtering
• Handles both single-language and batch processing modes

Why Use It

• Ensures final outputs comply with domain restrictions
• Maintains consistency between data files and their metadata
• Prepares clean data for subsequent deduplication steps

Usage

Configure domain_file path in config.yaml and run:

python result_filtering/final_domain_filter.py

Configuration

Uses these key config parameters:

domain_file: "path/to/domain_filter.txt"  # List of domains to exclude
output:
  directory: "output"                    # Where to find/save files

Output

Updates both:

[LANGUAGE]_filtered.json - With domain-filtered results

meta_data/[LANGUAGE]_meta_data.json - With filtered statistics

Step 4.2: Formatting Output for GlotWeb

Purpose

Transforms crawled language data into a structured format suitable for GlotWeb visualization, enriching it with metadata and linguistic information.

Key Features

• Extracts language metadata (speaker counts, language family)
• Checks inclusion in major multilingual datasets (MADLAD-400, Flores, Glot500)
• Organizes URLs by domain with site categorization
• Handles both single-language and batch processing

Why Use It

• Creates standardized format for GlotWeb frontend
• Enriches raw data with valuable linguistic metadata
• Provides domain-level organization of web resources
• Generates compatibility flags for popular multilingual datasets

Configuration

output:
  formated_directory: "formatted_output"  # Output directory
  formated_file_name: "{language}_formatted.json"  # Output filename pattern

Usage

python result_filtering/format_for_glotweb.py

Step 4.3: Robots.txt Compliance Filtering

Purpose

Filters out domains that explicitly block Common Crawl's CCBot in their robots.txt file, ensuring compliance with website crawling policies.

Key Features

• Checks each domain's robots.txt for CCBot restrictions
• Removes entire domains if they block CCBot
• Preserves all other metadata while filtering
• Handles both single-language and batch processing

Why Use It

• Ensures ethical web scraping compliance
• Prevents potential legal issues
• Maintains good web citizenship by respecting robots.txt
• Filters before final dataset compilation

Configuration

output:
  formated_directory: "formatted_output"  # Input directory (from Step 4.2)
  cleaned_directory: "cleaned_output"    # Output directory for filtered data

Usage

python result_filtering/robots_compliance_filter.py

Process Flow

• Loads formatted JSON from Step 4.2
• For each domain:
• Fetches robots.txt
• Checks for CCBot restrictions
• Saves cleaned version with compliant domains only

Output

• Maintains same structure as input
• Only contains domains that allow CCBot
• Saved as [LANGUAGE].json in cleaned directory

Notes

• If robots.txt is inaccessible, assumes crawling is allowed
• Only checks for explicit CCBot blocks (not general User-agent: *)
• Processes domains sequentially with 5-second timeout
• Preserves all non-URL metadata (speaker counts, language family etc.)

Step 4.4: Final Cleaning and Deduplication

Purpose

Performs final data cleaning through URL normalization and deduplication to create a polished dataset.

Process Overview

1. HTTP/HTTPS Merging (http_merge_2.py):

   • Combines duplicate sites with different protocols (http/https)
   • Standardizes www/non-www variants
   • Preserves all unique links

2. Hash Fragment Removal (remove_all_hash.py):

   • Removes URL fragments (#section)
   • Deduplicates URLs that only differ by fragments

Configuration

output:
  robots_filtered: "output/robots_filtered"  # Input from Step 4.3
  http_merged: "output/http_merged"         # Intermediate output
  deduplication: "output/deduplication"     # Final output

Usage

# Run protocol merging first
python result_filtering/http_merge_2.py

# Then run hash removal
python result_filtering/remove_all_hash.py

Key Features

• Protocol-agnostic site merging
• Consistent URL normalization
• Fragment removal while preserving query parameters
• Order-preserving deduplication

Output

Cleaned JSON files with:

• Unified site entries
• Normalized URLs
• No duplicate content

✅ Processed datasets in output/cleaned_output/[LANG].json containing:

• Verified web links
• Language metadata (speakers, family)
• Domain categorization
• Compatibility flags (FLORES/MADLAD/Glot500)

✅ Metadata reports in output/meta_data/ with:

• Crawling statistics
• Domain distributions
• Filtering metrics

Step 5: Dataset Validation & Community Contribution

Community Auditing Request

We urgently need native speakers and linguists to validate results:

How to Audit

1. Explore your language in the GlotWeb Demo
2. Check 10-20 random links for:
   • Actual language content (not machine translation)
   • Cultural/educational value
   • Flag as religious if content is from religious scriptures
   • Correct language/dialect labeling
3. Report issues via:
   • GitHub Issues
   • Email: in.sefat@tum.de

Why This Matters

Impact Area	Community Role
Data Quality	Remove spam/misclassified content
Language Preservation	Identify valuable resources
NLP Development	Improve training data for LLMs

Get Involved

• Speakers: Join us in Language auditing.
• Researchers: Use data with citation (BibTeX forthcoming)

Native speakers of underrepresented languages are especially needed!