GOT: Testing for Originality in Natural Language Generation

Jennifer Brooks, Abdou YoussefACL 2021

TL;DR

This paper introduces GOT (Generation Originality Test), a novel algorithm to automatically detect plagiarism in generated text by identifying fragments that are unique within a ground truth corpus. This is crucial for holding language generators accountable and protecting intellectual property in an era of increasingly sophisticated AI text generation.

What Problem Does This Paper Address?

The rapid advancement of Natural Language Generation (NLG) models, particularly those trained on vast datasets like the internet, raises significant ethical and legal questions about originality and plagiarism. While we can easily identify original ideas (like Einstein's E=mc²), it's much harder to determine if a generated use of language is original, especially when the generator has been trained on a massive corpus. Existing metrics like perplexity and BLEU are insufficient; high perplexity can indicate nonsense, and low BLEU might just mean a lack of overlap without guaranteeing originality. This paper tackles the challenge of automatically assessing whether a generated sentence represents an original use of language, distinct from simply regurgitating training data without attribution, and thus potentially infringing on intellectual property.

Methods

The core of the proposed solution is the Generation Originality Test (GOT). GOT operates by first identifying "original" fragments within a ground truth corpus. An "original" fragment is defined as a sequence of words (an n-gram) that appears exactly once in the entire ground truth. The algorithm systematically extracts all possible n-grams from the ground truth sentences, keeping track of their frequency.

To implement this, the authors propose using balanced binary search trees (like red-black trees) to store fragments and their counts efficiently. The process involves:

  1. Building a Fragment Database (F): Iterating through each sentence in the ground truth, extracting all n-grams of varying lengths (from 2 up to the sentence length). Each unique n-gram is added to a set F, and its count is maintained. The definition of "original" for ground-truth fragments is refined to consider occurrences across different documents or authors, incrementing the count only once per unique fragment instance. A fragment is considered "original" if its final count is 1.
  2. Testing Generated Sentences: For each generated sentence, GOT extracts its n-grams and checks if any of these fragments exist in the pre-computed F database. If a generated fragment is found in F and its count is 1 (meaning it's an "original" fragment from the ground truth), then the generated sentence is flagged as potentially violating copyright if no citation is provided. If the fragment appears more than once in the ground truth, it's considered "not original."

The algorithm's runtime complexity is analyzed as O(n log n), where n is the number of sentences in the ground truth, making it computationally feasible.

Datasets

The paper uses a custom dataset for their example application:

  • Ground Truth: A set of 22,113 quotes, each containing at least one metaphor and labeled with a topic. This dataset is derived from "Dr. Mardy's Dictionary of Metaphorical Quotations" (Grothe, 2008).
  • Training Data: The metaphor generator was trained on this ground truth dataset.
  • Generated Data: 500 metaphors were generated by their metaphor generator for evaluation.

Key Insights and Ideas

  • Distinguishing Language Use from Ideas: The paper makes a crucial distinction between generating an "original idea" and an "original use of language." GOT focuses on the latter, which is more directly testable with text analysis.
  • "Original" as Uniqueness: The core definition of originality for a fragment is its unique occurrence within a comprehensive ground truth. This provides a concrete, measurable criterion.
  • Plagiarism vs. Originality: GOT serves a dual purpose: identifying if a generated sentence is an original use of language, and flagging potential plagiarism by detecting the use of rare, potentially copyrightable fragments without attribution.
  • Scalability of the Approach: The O(n log n) complexity suggests that GOT can be applied to large corpora, making it practical for real-world NLG systems.

Key Results

In their experimental application to a metaphor generator, the authors generated 500 metaphors. Applying GOT to these generated metaphors revealed that:

  • Only 32 out of 500 generated metaphors contained an "original" fragment (a fragment appearing only once in the training set).
  • This implies that 468 metaphors (over 93%) were considered original uses of language by the GOT metric.

Limitations and Future Work

  • Ground Truth Representativeness: The effectiveness of GOT heavily relies on the ground truth corpus being sufficiently large and representative of the language space. If a truly original fragment is not present in the ground truth, GOT cannot detect it. The authors acknowledge that with very large datasets, the criterion for "original" might need to be relaxed from "once and only once" to a small number of occurrences.
  • Definition of "Fragment": The paper uses n-grams (sequences of tokens) as fragments. Future work could explore more sophisticated linguistic units or semantic chunks.
  • Dynamic Ground Truth: The authors suggest that the ground truth could ideally grow over time, incorporating newly generated and verified original sentences. This would require a mechanism for updating the fragment database.
  • Beyond Plagiarism: While GOT addresses plagiarism, it doesn't inherently measure the quality or creativity of the generated text beyond its originality. Future work could integrate GOT with other NLG evaluation metrics.

Summary updated

BibTeX
@inproceedings{brooks2021got,
  author = {Jennifer Brooks and Abdou Youssef},
  title = {{GOT: Testing for Originality in Natural Language Generation}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2021},
}
Figure