dneg

dneg is a Python package for detecting negation cues and their scopes in text using fine-tuned BERT models. It provides a pipeline to process batched text inputs, identify negation cues (e.g., "not", "n't"), and determine the scope of negation within sentences. The package leverages the Hugging Face Transformers library, PyTorch-Geometric, and PyTorch for efficient inference.

Features

• Negation Cue Detection: Identifies negation cues (e.g., "not", "n't") using the CueBertInference or CueBertInferenceGAT class.
• Negation Scope Detection: Determines the scope of negation in text using the ScopeBertInference or ScopeBertInferenceGAT class.
• Pipeline Processing: Combines cue and scope detection in a single pipeline for streamlined processing.
• Batch Processing: Supports batched inputs for efficient inference.
• GPU Support: Utilizes CUDA for accelerated inference on compatible hardware.
• TODO: In the future there will be negation event and focus detection components added to the Pipeline.
• 🌟✴️🌟 German Language Support: The pipeline now supports negation detection in German as well as English.
• 🌟✴️🌟 Multi Language Support: The pipeline now supports negation detection for 10 additional Languages: German, Italian, Spanish, French, Dutch, Chinese, Japanese, Russian, Hindi, Arabic
• We trained around 300 models, which can all accessed via this package (and can be found on huggingface)

Prerequisites

• Python 3.6 or higher
• PyTorch
• PyTorch Geometric
• Scikit-Learn
• UD-Pipe
• Spacy
• Hugging Face Transformers
• CUDA-enabled GPU (optional, for faster inference)

Install via PyPI

pip install dneg

Install Dependencies

Ensure dependencies are installed:

pip install torch transformers

Usage

Basic Example

The following example demonstrates how to use the Pipeline class to detect negation cues and scopes in a batch of sentences.

from dneg import PipelineTests, Pipeline
pipe = Pipeline.from_language()
batch_tokens = [
    "This is not an example for testing, it is also not an example for multi negation testing and i never ate spinach .".split(
        " "),
    ['In', 'contrast', 'to', 'anti-CD3/IL-2-activated', 'LN', 'cells', ',', 'adoptive', 'transfer', 'of',
     'freshly', 'isolated', 'tumor-draining', 'LN', 'T', 'cells', 'has', 'no', 'therapeutic', 'activity',
     '.'],
]
res = pipe.run(batch_tokens)
Pipeline.pretty_print(res)

Results in:

This                           S     X     X    
is                             S     X     X    
not                            C     X     X    
an                             S     X     X    
example                        X     X     X    
for                            S     X     X    
testing,                       X     X     X    
it                             X     S     X    
is                             X     S     X    
also                           X     S     X    
not                            X     C     X    
an                             X     S     X    
example                        X     X     X    
for                            X     S     X    
multi                          X     S     X    
negation                       X     X     X    
testing                        X     X     X    
and                            X     X     X    
i                              X     S     S    
never                          X     X     C    
ate                            X     X     S    
spinach                        X     X     S    
.                              X     X     X    

In                             X    
contrast                       X    
to                             X    
anti-CD3/IL-2-activated        X    
LN                             X    
cells                          X    
,                              X    
adoptive                       X    
transfer                       X    
of                             X    
freshly                        X    
isolated                       X    
tumor-draining                 X    
LN                             X    
T                              X    
cells                          X    
has                            X    
no                             C    
therapeutic                    X    
activity                       X    
.                              X  

Advanced Usage

For custom models or tokenizers, you can initialize the pipeline with specific components:

from dneg import Pipeline, CueBertInference, ScopeBertInference

# Load custom models and tokenizers
mcue_path = "D-NEG/cue-de-sfu"
mscope_path = "D-NEG/scope-de-sfu"

# Initialize pipeline with custom components
pipe = Pipeline(
    components=[CueBertInference, ScopeBertInference],
    model_paths=[mcue_path, mscope_path]
)

# Define input
batch_tokens = [
    "Das ist nicht ein Testsatz .".split(" ")
]

# Run inference
results = pipe.run(batch_tokens)

# Print results
Pipeline.pretty_print(results)

Package Structure

• CueBertInference: Detects negation cues (labeled as "C" for cues, "X" otherwise).
• ScopeBertInference: Identifies the scope of negation (labeled as "S" for scope, "X" otherwise).
• • CueBertInferenceGAT: Detects negation cues (labeled as "C" for cues, "X" otherwise) + syntax aware GAN.
• ScopeBertInferenceGAT: Identifies the scope of negation (labeled as "S" for scope, "X" otherwise) + syntax aware GAN.
• Pipeline: Combines CueBertInference and ScopeBertInference for end-to-end negation detection.
• Special Tokens:
  • [CUE]: Marks negation cues.
  • [SCO]: Marks negation scope.

Requirements

See requirements.txt for a full list of dependencies. Key dependencies include:

• torch>=1.9.0
• transformers>=4.9.0

License

This project is licensed under the MIT License - see the LICENSE file for details.