saadob12/t5_autochart_2

Tackling Hallucinations in Neural Chart Summarization

Introduction

The trained model for investigations and state-of-the-art (SOTA) improvements are detailed in the paper: Tackling Hallucinations in Neural Chart Summarization.

Abstract

Hallucinations in text generation occur when the system produces text that is not grounded in the input. In this work, we address the problem of hallucinations in neural chart summarization. Our analysis reveals that the target side of chart summarization training datasets often contains additional information, leading to hallucinations. We propose a natural language inference (NLI) based method to preprocess the training data and demonstrate through human evaluation that our approach significantly reduces hallucinations. Additionally, we found that shortening long-distance dependencies in the input sequence and adding chart-related information such as titles and legends enhances overall performance.

Main Findings from the Paper

• Enhanced Context Provision: Emphasizing the importance of providing more context and reducing long-distance dependencies in the input format.
• NLI Cleaning Step: Introducing an NLI-based cleaning step to eliminate ungrounded information in the training data.
• Reduction of Intrinsic Hallucinations: Demonstrating that reducing long-distance dependencies and adding more context leads to fewer intrinsic hallucinations.
• Cause of Extrinsic Hallucinations: Identifying that extrinsic hallucinations are caused by ungrounded information in training summaries.
• Human Evaluation Results: Showing that using NLI to filter training summaries significantly reduces hallucinations.

Training Data

GitHub Link for the Original AutoChart Data

AutoChart

Training Details

The model was trained using optimized prompts based on findings from the paper Tackling Hallucinations in Neural Chart Summarization.

• Optimized Prompt Dataset: Hallucinations-C2T Data

Example Use

Input Prompt Template

The input prompt template consists of the title, x-y labels, and x-y values.

Appending C2T Prefix: Before every input to the model, append C2T:.

    from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
    
    tokenizer = AutoTokenizer.from_pretrained("saadob12/t5_C2T_big")
    model = AutoModelForSeq2SeqLM.from_pretrained("saadob12/t5_C2T_big")
    
    data = 'Breakdown of coronavirus (COVID-19) deaths in South Korea as of March 16, 2020\n' \
        'by chronic disease x-y labels\n' \
        'Response - Share of cases, x-y values Circulatory system disease* 62.7%, ' \
        'Endocrine and metabolic diseases** 46.7%, Mental illness*** 25.3%, ' \
        'Respiratory diseases*** 24%, Urinary and genital diseases 14.7%, Cancer 13.3%, ' \
        'Nervous system diseases 4%, Digestive system diseases 2.7%, Blood and hematopoietic diseases 1.3%' \
    
    prefix = 'C2T: '
    tokens = tokenizer.encode(prefix + data, truncation=True, padding='max_length', return_tensors='pt')
    generated = model.generate(tokens, num_beams=4, max_length=256)
    tgt_text = tokenizer.decode(generated[0], skip_special_tokens=True, clean_up_tokenization_spaces=True)
    summary = tgt_text.strip('[]""')
    # Summary:
    #As of March 16, 2020, around 62.7% of COVID-19 deaths in South Korea were related to circulatory system diseases.
    #Other chronic diseases include endocrine and metabolic diseases, mental illness, and cancer.
    #South Korea confirmed 30,017 infection cases, including 501 deaths.

Intended Use and Limitations

You can use the model to generate summaries of data files. Works well for general statistics like the following:

Year	Children born per woman
2018	1.14
2017	1.45
2016	1.49
2015	1.54
2014	1.6
2013	1.65

May or may not generate an okay summary at best for the following kind of data:

Model	BLEU score	BLEURT
bert-small	25.4	-0.11
bert-base	28.2	0.12
bert-large	35.4	0.34

Citation

Kindly cite the work. Thank you.

    @inproceedings{obaid-ul-islam-etal-2023-tackling,
        title = {Tackling Hallucinations in Neural Chart Summarization},
        author = {Obaid ul Islam, Saad and Škrjanec, Iza and Dusek, Ondrej and Demberg, Vera},
        booktitle = {Proceedings of the 16th International Natural Language Generation Conference},
        month = sep,
        year = {2023},
        address = {Prague, Czechia},
        publisher = {Association for Computational Linguistics},
        url = {https://aclanthology.org/2023.inlg-main.30},
        doi = {10.18653/v1/2023.inlg-main.30},
        pages = {414--423},
        abstract = {Hallucinations in text generation occur when the system produces text that is not grounded in the input. In this work, we tackle the problem of hallucinations in neural chart summarization. Our analysis shows that the target side of chart summarization training datasets often contains additional information, leading to hallucinations. We propose a natural language inference (NLI) based method to preprocess the training data and show through human evaluation that our method significantly reduces hallucinations. We also found that shortening long-distance dependencies in the input sequence and adding chart-related information like title and legends improves the overall performance.}
    }