Abs-package 0.0.1

A Streamlit-based application to compute Harmonic Abstraction and Rouge Complement Abstraction a formula based approch

Abstraction Metrics Calculator

This package provides a Streamlit-based application to compute Harmonic Abstraction and Rouge Complement Abstraction metrics for text summarization tasks. It processes JSON files containing documents, reference summaries, and model-generated summaries, and returns detailed metric analysis including ROUGE complements and various abstraction ratios.

Features

• Calculate Harmonic Abstraction based on the input text, reference summary, and model-generated summary.
• Compute ROUGE Complement Abstraction for ROUGE-1, ROUGE-2, and ROUGE-L scores.
• Calculate Reference Abstraction (RA), Summary Abstraction (SA), and Abstraction Ratio (AR).
• Calculate additional metrics such as Length Ratios and Weighted Overlaps.
• Process multiple entries from a JSON file.
• Generate downloadable CSV reports for further analysis.

Installation

1. Install the required dependencies using the following commands:

   pip install streamlit
   pip install pandas
   pip install rouge
   

Usage

1. Run the Streamlit app with the following command:

   streamlit run app.py
   

2. Upload a JSON file in the format:

   [
     {
       "Index": 0,
       "Document": "The input text for the summarization task...",
       "Summary": "The reference summary...",
       "Models": {
         "model_name_1": "Generated summary by model 1",
         "model_name_2": "Generated summary by model 2"
       }
     }
   ]
   

Usage

Once the package is installed, you can access all the functionalities it provides. Below is a guide to the available functions and how to use them.

To use the functions provided in this package, you need to import them from the respective modules based on your use case (single text processing or large text processing). Here’s a guide on how to access each functionality.

single text processing

from Abs_harmonic_single.Abs_harmonic import ( calculate_ra, calculate_sa, calculate_ar, abs_Harmonic, abs_rouge_complement )

input_text = "This is the original text." reference = "This is a summary of the original text." generated_summary = "This is the generated summary."

Calculate Harmonic Abstraction

harmonic_result = abs_Harmonic(input_text, reference, generated_summary) print(f"Harmonic Abstraction: {harmonic_result}")

Calculate Rouge Complement Abstraction

ra_complement, sa_complement = abs_rouge_complement(input_text, reference, generated_summary) print(f"Rouge Complement Abstraction (RA): {ra_complement}") print(f"Rouge Complement Abstraction (SA): {sa_complement}")

In Abs_harmonic_single.Abs_harmonic

1. tokenize_text(text) Tokenizes the input text into individual words. Parameters: text (str): The input text. Returns: A list of words in the text.
2. calculate_overlap(text1, text2) Calculates the overlap (intersection) of unique words between two pieces of text. Parameters: text1 (str): The first text (e.g., input text). text2 (str): The second text (e.g., reference or generated summary). Returns: The count of common words between the two texts.
3. calculate_ra(reference, input_text, overlap) Calculates the Reference Abstraction (RA) score. Parameters: reference (str): The reference summary. input_text (str): The input text. overlap (int): The overlap count between reference and input_text. Returns: The calculated RA score.
4. calculate_sa(summary, input_text, overlap) Calculates the Summary Abstraction (SA) score. Parameters: summary (str): The generated summary. input_text (str): The input text. overlap (int): The overlap count between summary and input_text. Returns: The calculated SA score.
5. calculate_ar(ra, sa) Calculates the Abstraction Ratio (AR) as the ratio of SA to RA. Parameters: ra (float): The Reference Abstraction (RA) value. sa (float): The Summary Abstraction (SA) value. Returns: The AR value.
6. abs_Harmonic(input_text, reference, generated_summary) Calculates the Harmonic Abstraction score using the RA and SA values derived from the input text, reference summary, and generated summary. Parameters: input_text (str): The original input text. reference (str): The reference summary. generated_summary (str): The generated summary. Returns: The harmonic abstraction score.
7. abs_rouge_complement(input_text, reference, generated_summary) Calculates the Rouge Complement Abstraction by combining the RA, SA values and the Rouge Complement score. Parameters: input_text (str): The original input text. reference (str): The reference summary. generated_summary (str): The generated summary. Returns: A tuple containing the Rouge Complement Abstraction values for RA and SA.

For large data sets: (Access the below package)

from Abs_harmonic_large_process.Abs_harmonic_large_v2

8. Click Process Data to calculate Harmonic Abstraction, Rouge Complement Abstraction, and other metrics for each model's generated summary.

9. Download the processed results in CSV format.

Metrics

• Harmonic Abstraction (RA, SA, AR):

  • Measures abstraction based on overlap between the input text, reference summary, and model-generated summary.
  • Calculates Reference Abstraction (RA), Summary Abstraction (SA), and Abstraction Ratio (AR).

• Rouge Complement Abstraction:

  • Complements of ROUGE-1, ROUGE-2, and ROUGE-L scores for each model-generated summary.

• Length Ratios:

  • Normalized ratio of lengths between input text and reference summary, as well as input text and generated summary.

• Weighted Overlap:

  • Frequency-based overlap between tokens in the input text, reference summary, and generated summary.

Example

After running the application, you can upload a JSON file, calculate metrics, and download the results as a CSV file.

The output file will include metrics such as:

• Harmonic Abstraction (RA, SA, AR)
• Rouge1 Complement RA/SA/AR
• Rouge2 Complement RA/SA/AR
• RougeL Complement RA/SA/AR
• Length Ratios and Weighted Overlaps

To access our research paper please click below:

https://drive.google.com/file/d/1_tnkzaMYOuFWWf31OdYjldfZG1XSdkcb/view?usp=sharing

License

This project is licensed under the MIT License. See the LICENSE file for more details.