Does a Large Language Model Really Speak in Human-Like Language?

TL;DR

The paper investigates whether large language models speak in human-like language by analyzing latent community structures in embeddings of human-authored text ($\mathcal{O}$) and LLM-generated paraphrases ($\mathcal{G}$, $\mathcal{G}'$, $\mathcal{S}$). It introduces an anchor-based, paired hypothesis-testing framework, using clustering on embeddings and mapped distances $\mathbb{D}_j$ to compare distributions via Johnson's paired modified t-test with permutation. The study encompasses multi-source data, five temperature settings, and various cluster counts, demonstrating that GPT-generated text remains structurally distinct from human-written text across settings, with nuanced behavior as paraphrase depth and generation variability change. The authors contribute a statistically principled method for cross-dataset comparison in NLP and provide empirical evidence challenging the notion that current LLM paraphrasing fully matches human language in latent structure.

Abstract

Large Language Models (LLMs) have recently emerged, attracting considerable attention due to their ability to generate highly natural, human-like text. This study compares the latent community structures of LLM-generated text and human-written text within a hypothesis testing procedure. Specifically, we analyze three text sets: original human-written texts ($\mathcal{O}$), their LLM-paraphrased versions ($\mathcal{G}$), and a twice-paraphrased set ($\mathcal{S}$) derived from $\mathcal{G}$. Our analysis addresses two key questions: (1) Is the difference in latent community structures between $\mathcal{O}$ and $\mathcal{G}$ the same as that between $\mathcal{G}$ and $\mathcal{S}$? (2) Does $\mathcal{G}$ become more similar to $\mathcal{O}$ as the LLM parameter controlling text variability is adjusted? The first question is based on the assumption that if LLM-generated text truly resembles human language, then the gap between the pair ($\mathcal{O}$, $\mathcal{G}$) should be similar to that between the pair ($\mathcal{G}$, $\mathcal{S}$), as both pairs consist of an original text and its paraphrase. The second question examines whether the degree of similarity between LLM-generated and human text varies with changes in the breadth of text generation. To address these questions, we propose a statistical hypothesis testing framework that leverages the fact that each text has corresponding parts across all datasets due to their paraphrasing relationship. This relationship enables the mapping of one dataset's relative position to another, allowing two datasets to be mapped to a third dataset. As a result, both mapped datasets can be quantified with respect to the space characterized by the third dataset, facilitating a direct comparison between them. Our results indicate that GPT-generated text remains distinct from human-authored text.

Figures (7)

• Figure 1: Text generation framework
• Figure 2: User review text example
• Figure 3: An example under which the latent community structure of $\mathbf{D}_1$ and $\mathbf{D}_2$ are identical. The points labeled a, b, and c appearing in all plots indicate the paired points. Across all plots, clusters are color-coded. Top row: the original forms of $\mathbf{A}$, $\mathbf{D}_1$, and $\mathbf{D}_2$. Bottom row: the color coding for the communities of $\mathbf{D}_1$ and $\mathbf{D}_2$ mapped onto $\mathbf{A}$ from left to right, along with the corresponding community centers, marked with crosses.
• Figure 4: An example under which the latent community structure of $\mathbf{D}_1$ and $\mathbf{D}_2$ are distinct. The points labeled a, and b, and c appearing in all plots indicate the paired points. Across all plots, clusters are color-coded. Top row: the original forms of $\mathbf{A}$, $\mathbf{D}_1$, and $\mathbf{D}_2$. Bottom row: the color coding for $\mathbf{D}_1$ and $\mathbf{D}_2$ mapped onto $\mathbf{A}$ from left to right, along with the corresponding community centers, marked with crosses.
• Figure 5: Kullback-Leibler Divergence and Wasserstein Distance between $\mathbb{D}_1$ and $\mathbb{D}_2$ versus the temperature parameter $\rho$ for different values of $K$ ($K = 2, 3, 4, 5$) in the Review data. The settings used to calculate $\mathbb{D}_1$ and $\mathbb{D}_2$ in $(\mathcal{O}, \mathcal{S})$, $(\mathcal{O}, \mathcal{G'})$, $(\mathcal{G}, \mathcal{S})$, and $(\mathcal{G}, \mathcal{G'})$ correspond to those used in testing $H_0 (\mathcal{G}, \{\mathcal{O}, \mathcal{S}\})$, $H_0 (\mathcal{G}, \{\mathcal{O}, \mathcal{G'}\})$, $H_0 (\mathcal{O}, \{\mathcal{G}, \mathcal{S}\})$, and $H_0 (\mathcal{O}, \{\mathcal{G}, \mathcal{G'}\})$, respectively. Each row corresponds to a specific $K$ value: (Top to Bottom) $K = 2$, $K = 3$, $K = 4$, and $K = 5$. Within each row, the left plot represents Kullback-Leibler Divergence, and the right plot represents Wasserstein Distance.

Theorems & Definitions (2)

• Remark 1
• Remark 2