Portparser v2

A parsing model for Brazilian Portuguese following the Universal Dependencies (UD) framework.

Note: This is a fork of the original Portparser.v2 Hugging Face Space by NILC-ICMC-USP. The original project includes a Streamlit web interface (app.py) which is still present in this repository but is not required for using the parser programmatically.

Installation

# Using uv (recommended)
uv pip install -e .

# Or using pip
pip install -e .

Optional dependencies

# For the Streamlit UI (not required for API usage)
uv pip install -e ".[ui]"

# For development/testing
uv pip install -e ".[dev]"

Usage

Quick Start

The simplest way to parse text:

from portparser_v2 import parse

# Parse raw text (with automatic sentence segmentation)
conllu_output = parse("O Brasil é um país tropical. Ele fica na América do Sul.")
print(conllu_output)

API Reference

parse(text, segment=True) -> str

Quick parse function that returns CoNLL-U content directly.

from portparser_v2 import parse

# With sentence segmentation (default)
result = parse("O Brasil é um país tropical.")

# Without segmentation (assumes one sentence per line)
result = parse("Primeira frase.\nSegunda frase.", segment=False)

Parameters:

• text: Input text to parse
• segment: Whether to automatically segment sentences (default: True)

Returns: Parsed CoNLL-U content as a string.

---

parse_text(text, output_path=None, work_dir=None, model_path=None, segment_sentences=False) -> str

Full-featured parsing function with more control over the pipeline.

from portparser_v2 import parse_text

# Basic usage
output_file = parse_text("O Brasil é um país tropical.", segment_sentences=True)

# With custom output path
output_file = parse_text(
    "O Brasil é um país tropical.",
    output_path="./output.conllu",
    work_dir="./temp",
    segment_sentences=True
)

Parameters:

• text: Input text to parse
• output_path: Optional path for final CoNLL-U output (uses temp file if None)
• work_dir: Optional working directory for temp files (creates temp dir if None)
• model_path: Optional path to model weights (downloads from HuggingFace if None)
• segment_sentences: If True, run sentence segmentation first; if False, assume one sentence per line

Returns: Path to the final parsed CoNLL-U file.

---

parse_file(input_path, output_path=None, work_dir=None, model_path=None, segment_sentences=False) -> str

Parse text from a file.

from portparser_v2 import parse_file

# Parse a text file
output_file = parse_file(
    "input.txt",
    output_path="output.conllu",
    segment_sentences=True
)

Parameters: Same as parse_text, but with input_path instead of text.

Returns: Path to the final parsed CoNLL-U file.

Pipeline

The parser runs a 4-step pipeline:

1. Sentence Segmentation (optional) - splits raw text into sentences
2. Tokenization - converts sentences to CoNLL-U format with tokens
3. Parsing/Prediction - runs the LatinPipe neural model to predict POS tags, morphological features, lemmas, and dependency relations
4. Post-processing - applies cleanup and corrections to the output

Model

The model weights are automatically downloaded from HuggingFace on first use:

• Repository: lucelene/Portparser.v2-latinpipe-core

License

See LICENSE for details.