Divergent strategies in text simplification: A comparative analysis of AI and human approaches in language processing

1 Introduction

The increasing importance of text simplification in recent years is driven by factors such as rising global migration and the growing need for accessible information (Iori 2016; Sato 2004). Initiatives like “Easy Japanese” (Yasashii Nihongo) exemplify this trend, aiming to make complex Japanese texts more accessible to non-native speakers and others who may find standard Japanese challenging (Sato 1996). Simultaneously, the emergence of large language models (LLMs) has dramatically expanded the potential for automating text simplification processes, creating new possibilities for addressing accessibility challenges at scale. This technological advancement, underpinned by developments such as transformer architectures (Vaswani et al. 2017) and the capabilities demonstrated by models like GPT-3 (Brown et al. 2020), raises crucial questions about how AI-generated simplifications compare to those produced by humans and what these differences might reveal about language processing more broadly.

This study aims to deepen our understanding of text simplification processes by directly comparing AI-generated and human-produced simplifications of Japanese news articles. Specifically, this research addresses the following research questions:

1. Can AI models like GPT-4 outperform human experts in Japanese text simplification tasks as measured by readability metrics and content preservation?

2. What are the specific strategies employed by AI and humans when simplifying Japanese texts, and how do these strategies differ?

3. What implications do these differences have for our understanding of human cognitive processes in language comprehension and production?

To address these questions, we analyze a corpus of 420 original Japanese news articles alongside their simplified versions created by professional editors at NHK’s “NEWS WEB EASY” initiative and those generated by ChatGPT (GPT-4).^[1] Our analysis employs a readability formula specifically designed for Japanese texts (Lee and Hasebe 2020) and a detailed examination of various textual features, including sentence length, vocabulary complexity, and information retention. This comparative approach is intended to not only illuminate the strengths (cf. Wei et al. 2022) and limitations (cf. Bender and Koller 2020) of AI in language processing but also offer a perspective through which we can gain a deeper understanding of the cognitive mechanisms underlying human language use, particularly in the context of simplification tasks.

The practical significance of this research is considerable, particularly for initiatives like NHK’s “NEWS WEB EASY.” Relying solely on human editors for creating simplified content presents challenges, including limitations in production volume and potential inconsistencies or subjectivity inherent in the manual editing process (Tanaka et al. 2018). Generative AI offers the potential to augment these efforts, assisting in the expansion of accessible Japanese content. This could allow human editors to concentrate on tasks requiring nuanced judgment and expertise, ultimately supporting broader accessibility goals for diverse language users.

While text simplification has a long history in computational linguistics (e.g., Shardlow 2014; Siddharthan 2014), direct comparisons between AI and human approaches remain relatively scarce, especially for languages with unique characteristics like Japanese. Our methodology combines quantitative assessment of readability with detailed analysis of linguistic features, providing a comprehensive picture of the distinct strategies employed by humans and AI. The insights gained from this comparison suggest promising possibilities in collaborative human-AI approaches to text simplification – leveraging the systematic transformation capabilities of AI alongside human expertise in cultural contextualization and pedagogical adaptation to create more effective and accessible content for language learners and non-native speakers.

The remainder of this paper proceeds as follows: Section 2 reviews relevant literature on text simplification, “Easy Japanese,” readability assessment, and AI advances in natural language processing. Section 3 details our methodology and dataset, while Section 4 presents our quantitative and qualitative findings. Section 5 discusses the implications of these results, and Section 6 concludes with a summary of key findings and their broader implications.

2 Background and related work

Text simplification, a research area within computational linguistics and natural language processing, has been actively explored for several decades. As defined by Shardlow (2014), text simplification is the process of reducing linguistic complexity while preserving the original meaning and information. Its goal is to enhance text accessibility for diverse groups, including second language learners, individuals with reading difficulties, and those seeking simplified access to complex information. The field has progressively shifted from rule-based systems employing handcrafted rules (Devlin and Tait 1998) towards more sophisticated data-driven methods, incorporating machine learning techniques and ultimately leading to the development of large language models capable of generating more context-aware and fluent simplifications (Alva-Manchego et al. 2020; Siddharthan 2014). This evolution has opened up new possibilities for exploring how both humans and AI systems approach the complex task of text simplification.

3 Methodology

This study employs a comparative analysis to investigate the differences between AI-generated and human-produced text simplifications in Japanese, focusing on news articles as a representative genre. Our dataset consists of 420 original news articles sourced from NHK (Japan Broadcasting Corporation), along with their corresponding simplified versions created by both human experts and the AI model ChatGPT (GPT-4).^[2] dataset builds upon the work of Tanaka and Lee (2017), who compiled a collection of general news articles and their simplified counterparts from NHK’s “NEWS WEB EASY” (Tanaka et al. 2018). We expanded this resource by utilizing ChatGPT to generate AI-simplified versions of the same original articles, creating a three-way parallel corpus (Original – Human Simplified – AI Simplified). The news articles cover a diverse range of topics, including society, international affairs, economy, sports, and politics, as detailed in Table 1. This topical diversity allows our analysis to assess the performance of both human experts and AI across various subject matters and levels of linguistic complexity, providing a more robust and nuanced understanding of their respective simplification strategies.

The human-simplified texts in the dataset represent authentic examples of “Easy Japanese,” initially proposed by Sato (1996). As documented in Tanaka et al. (2018), each simplified text is produced through collaboration between Japanese language teachers and professional reporters at NHK who jointly rewrite regular news content. Their simplification process follows principles detailed in guidelines like those published by the Ministry of Internal Affairs and Communications, with the aim of enhancing text accessibility for non-native Japanese speakers with limited language proficiency (Iori 2016). The simplified texts generally correspond to JLPT N4 or N5 levels, representing basic Japanese proficiency.^[3] This involves utilizing simple vocabulary and grammar, constructing shorter sentences, minimizing the use of loanwords, avoiding idiomatic expressions and passive voice, and employing consistent sentence-final expressions.

To generate the AI-simplified versions of the news articles, we used ChatGPT (GPT-4 version, accessed May 2023), a large language model capable of performing complex text transformations. Instead of providing ChatGPT with detailed linguistic simplification guidelines (e.g., specific vocabulary lists or grammatical constraints), we opted for a more goal-oriented instruction. We prompted the AI to simplify the text for non-native Japanese speakers with limited language proficiency, targeting a level equivalent to JLPT N4 or N5. This approach, drawing on preliminary tests and findings by Lee (2023), was intended to explore the AI’s inherent simplification capabilities more naturally, as overly specific instructions can sometimes constrain the model’s performance and lead to suboptimal results (cf. Wei et al. 2022). The prompt used for the AI simplification task is presented below:

PROMPT: The following text is a news article broadcast by NHK. This news was created for native Japanese speakers, so it may be difficult for foreigners with limited Japanese language proficiency to understand. Please rewrite this in simple Japanese that can be understood by foreigners with limited Japanese language proficiency. The definition of simple Japanese is as follows: Please set it to match JLPT N5 or N4 level. N5 level is the level at which one can understand basic Japanese to some extent. Specifically, it is the level at which one can read and understand typical phrases and sentences written in hiragana, katakana, and basic kanji used in daily life. N4 level is the level at which one can understand basic Japanese. Specifically, it is the level at which one can read and understand passages on familiar everyday topics written using basic vocabulary and kanji.

- - -

[Original NEWS TEXT inserted here]

To quantitatively assess the readability of the original and simplified texts, we utilized the Japanese readability formula developed by Lee and Hasebe (2020), as introduced in Section 2.2. This formula is implemented within the jReadability system (see Tools section). It calculates a readability score (y) based on several linguistic features, expressed as follows:

Where the ratios typically represent the percentage occurrence of each word type or part-of-speech within the text. A higher score indicates lower complexity (i.e., easier readability).

The resulting readability score ranges from 0.5 (most difficult) to 6.4 (easiest), corresponding to different levels of Japanese proficiency, from advanced (0.5–1.4) to beginner (5.5–6.4). Detailed guidelines for interpreting these readability scores can be found in the Appendix. Previous research has demonstrated that this formula correlates strongly with human judgments of text difficulty, and it is widely adopted in Japanese language education research, making it a reliable and valid measure for our study (Hasebe and Lee 2015; Lee and Hasebe 2020). By utilizing this readability formula, we can objectively compare the effectiveness of human and AI simplification strategies in reducing text complexity and enhancing accessibility for the target audience.

To gain a deeper understanding of the specific strategies employed by human experts and AI in simplifying Japanese news articles, we conducted a detailed analysis of various textual features, complementing the readability scores obtained using the Lee and Hasebe (2020) formula. This analysis focused on features known to influence text complexity and readability, encompassing:

1. Sentence-level characteristics: total number of sentences, total number of morphemes, and average sentence length.

2. Lexical characteristics: rate of common nouns, rate of intermediate-level vocabulary, and rate of kango (Sino-Japanese words).

To extract these features, we first performed morphological analysis using MeCab, a widely-used tokenizer for Japanese text (see Tools section).^[4] Based on this analysis, we calculated the frequency of each feature within each text type (original, human-simplified, and AI-simplified). For the identification of intermediate-level vocabulary, we consulted the vocabulary lists provided by the Japanese Language Proficiency Test (JLPT) (Japan Foundation and Japan Educational Exchanges and Services 2004). By examining these textual features, we aim to go beyond the overall readability scores and gain deeper insights into the specific linguistic choices made by humans and AI during the simplification process.

To determine whether statistically significant differences existed among the original texts, human-simplified versions, and AI-simplified versions in terms of readability and the previously described textual features, we performed a series of statistical analyses. Specifically, we employed one-way analysis of variance (ANOVA) for the readability score and each of the textual features outlined above. ANOVA is a statistical test used to compare the means of three or more groups and identify whether any overall significant difference exists among them. When the ANOVA revealed a significant difference among the groups (i.e., p < 0.001), we then performed post-hoc Tukey’s Honest Significant Difference (HSD) tests to determine which specific pairs of groups (original vs. human, original vs. AI, human vs. AI) differed significantly from each other. This two-step approach allowed us to identify both the presence and the location of significant differences in readability and textual features across the three text versions. All statistical analyses were performed using SPSS version 29, with the significance level set at p < 0.001.

While our quantitative analyses provide valuable insights into readability and textual features, we also recognized the importance of qualitative assessment to evaluate the faithfulness of the AI-generated simplifications. To this end, we performed a manual review of all 420 AI-simplified news articles. The primary focus of this review was to verify the preservation of essential information from the original articles. Specifically, we examined each AI-simplified text to ensure it accurately conveyed the core 5W1H elements (who, what, when, where, why, and how) of the corresponding original news story. This meticulous process allowed us to assess the AI’s ability to maintain factual accuracy and avoid introducing significant distortions or omissions during simplification.

In summary, our methodology adopts a multi-faceted approach, integrating quantitative analysis of readability scores and textual features with qualitative assessment of content preservation faithfulness. This comprehensive framework allows us to thoroughly compare the performance of human experts and AI in simplifying Japanese news articles, providing a detailed and nuanced understanding of their respective strengths, limitations, and simplification strategies.

4 Results and analysis

Our analysis of readability scores revealed significant differences among the original news articles (C1), human-simplified versions (C2), and AI-simplified versions (C3), indicating the effectiveness of both human and AI simplification methods. A one-way analysis of variance (ANOVA) confirmed a statistically significant difference in readability scores across the three corpora (F(2,1257) = 1,321.24, p < 0.001, η² = 0.678). This large effect size (η² = 0.678) indicates that the type of corpus (original, human, or AI) accounts for a substantial proportion of the variance in readability scores. Post-hoc Tukey’s HSD tests further revealed significant differences between all pairs of corpora (p < 0.001), confirming that each corpus type represents a distinct level of readability. The mean readability scores were lowest for the original articles (C1: M = 1.31, SD = 0.79), followed by the human-simplified versions (C2: M = 2.93, SD = 0.60), with the AI-simplified versions (C3: M = 3.69, SD = 0.63) achieving the highest scores. These findings suggest that while both human experts and AI successfully simplified the original texts, the AI-simplified versions demonstrated significantly higher readability on average, indicating a greater degree of simplification. Figure 1 visually illustrates these differences in the distribution of readability scores across the three corpora, showing a clear shift towards higher readability from C1 to C2, and further to C3.

Figure 1:

Distribution of readability scores.

To further visualize the distribution of readability scores, Figure 2 presents box plots for each corpus, illustrating the median, quartiles, and outliers. Notably, while the original articles (C1) and human-simplified versions (C2) show several outliers, the AI-simplified versions (C3) have no outliers, suggesting high consistency in the AI’s simplification process across different text types. As depicted in Figure 2, the AI-simplified versions exhibit the highest median readability score (3.66), falling within the “average” difficulty range according to our readability scale (see Appendix). The human-simplified versions show a lower median score (2.97), categorized as “somewhat difficult,” while the original articles have the lowest median score (1.33), corresponding to the “very difficult” level. The interquartile range (IQR) is largest for the original articles, indicating greater variability in the difficulty of original texts, while the IQRs for both simplified versions are smaller. The absence of outliers in the AI-simplified versions suggests a potential advantage over human simplification, where individual variation and subjective judgments can lead to greater inconsistency in the simplified texts.

Figure 2:

Box plots showing the distribution of readability scores.

To investigate whether the effectiveness of simplification varies across different genres of news articles, we examined the distribution of readability scores for each genre using box plots, as shown in Figure 3. The figure reveals distinct patterns in readability across the three corpora (C1: original, C2: human-simplified, C3: AI-simplified) and five genres (Economy, International, Politics, Society, and Sports). In the original corpus (C1, green), sports articles exhibited higher readability scores compared to other genres, with a median score of 1.97. This trend was maintained in both the human-simplified (C2, pink) and AI-simplified (C3, blue) versions, although readability scores were generally higher across all genres in the simplified versions. For instance, the median readability score for sports articles increased to 3.40 in the human-simplified versions and 4.69 in the AI-simplified versions. Notably, the AI-simplified versions (C3) demonstrated a more pronounced improvement in readability for sports articles compared to the human-simplified versions, with a difference in median scores of 1.29 (4.69 – 3.40). This suggests that the AI may be more adept at capturing and enhancing the inherent readability of sports articles, potentially due to their simpler sentence structures and more common vocabulary.

Figure 3:

Box plots of the distribution of readability scores across different news genres.

Furthermore, across all genres, the AI-simplified versions consistently showed higher median readability scores and smaller interquartile ranges compared to the human-simplified versions. For example, in the “Society” genre, the median readability score for AI-simplified versions was 3.54, while it was 2.85 for human-simplified versions. This pattern indicates that AI not only achieved higher overall readability but also demonstrated greater consistency in the simplification process across different genres, potentially leading to more predictable and reliable outcomes for various types of news content.

To gain a deeper understanding of the specific simplification strategies employed by humans and AI, we first examined sentence-level characteristics by analyzing various textual features. Figure 4 presents the mean total sentence count across the three corpora (C1: original, C2: human-simplified, C3: AI-simplified). A one-way ANOVA revealed statistically significant differences in sentence count among the three groups (F(2,1257) = 164.56, p < 0.001, η² = 0.245). This indicates that the type of simplification significantly affected the number of sentences in the resulting text. Post-hoc Tukey’s HSD tests further confirmed that all pairwise comparisons were significant (p < 0.001), suggesting that each simplification method resulted in a distinct sentence count.

Figure 4:

Mean total sentence count.

As shown in Figure 4, the AI-simplified versions (C3) contained the highest number of sentences on average (M = 16.10), followed by the original articles (M = 13.00), with the human-simplified versions (C2) having the lowest sentence count (M = 7.51). This pattern suggests that the AI tended to break down longer, more complex sentences into multiple shorter sentences, potentially increasing readability by reducing sentence complexity. In contrast, human simplifiers appeared to opt for a more extensive reduction of content, resulting in fewer sentences overall. This difference in approach highlights a key distinction between AI and human simplification strategies: AI appears to prioritize maintaining information while reducing sentence complexity, whereas humans may prioritize conciseness and brevity, potentially at the expense of some information loss.

Continuing our exploration of sentence-level characteristics, we examined the total morpheme count and mean sentence length across the three corpora. Figures 5 and 6 illustrate these measures, respectively. For both total morpheme count and mean sentence length, one-way ANOVA tests revealed statistically significant differences among the three groups (morpheme count: F(2,1257) = 231.18, p < 0.001, η² = 0.307; sentence length: F(2,1257) = 2042.41, p < 0.001, η² = 0.781). These findings indicate that the simplification method significantly impacted both the amount of content (as reflected in morpheme count) and the length of sentences. Post-hoc Tukey’s HSD tests further confirmed that all pairwise comparisons were significant (p < 0.001), suggesting that each simplification method resulted in distinct values for both measures.

Figure 5:

Mean total morpheme count.

Figure 6:

Mean sentence length.

As depicted in Figure 5, the original articles (C1) had the highest mean total morpheme count (M = 598), followed by the AI-simplified versions (C3, M = 332), with the human-simplified versions (C2, M = 205) having the lowest count. Conversely, as shown in Figure 6, the original articles (C1) had the longest mean sentence length (M = 47.73), followed by the human-simplified versions (C2, M = 27.75), with the AI-simplified versions (C3, M = 21.15) having the shortest sentences. These findings align with our previous observation that AI tended to break down longer sentences into multiple shorter ones while preserving more of the original content, as evidenced by the higher morpheme count in AI-simplified versions compared to human-simplified versions. In contrast, human simplifiers appeared to prioritize reducing both overall content and sentence complexity, resulting in fewer morphemes and shorter sentences, albeit longer than those produced by the AI. This difference in approach further underscores the distinct strategies employed by humans and AI in achieving text simplification.

Shifting our focus from sentence-level features to lexical characteristics, we examined the distribution of the rate of common nouns, the intermediate-level vocabulary usage rate, and the kango (Chinese-origin words) usage rate across the three corpora.^[5] Figures 7–9 visually represent these measures, respectively. One-way ANOVA tests revealed statistically significant differences among the three groups for all three lexical features (common nouns: F(2,1257) = 302.74, p < 0.001, η² = 0.325; intermediate-level vocabulary: F(2,1257) = 875.03, p < 0.001, η² = 0.582; kango: F(2,1257) = 297.57, p < 0.001, η² = 0.325). This indicates that the simplification method significantly influenced the lexical composition of the texts. Post-hoc Tukey’s HSD tests further confirmed that all pairwise comparisons were significant (p < 0.001), suggesting that each simplification method resulted in distinct lexical profiles.

Figure 7:

Mean common nouns count.

Figure 8:

Mean rate of intermediate-level words.

Figure 9:

Mean rate of Sino-Japanese words.

As depicted in Figure 7, the original articles (C1) exhibited the highest mean rate of common nouns (M = 167), followed by the AI-simplified versions (C3, M = 71), with the human-simplified versions (C2, M = 46) having the lowest rate. Figure 8 shows that the original articles (C1) also had the highest rate of intermediate-level vocabulary usage (M = 0.27), followed by the AI-simplified versions (C3, M = 0.14), and the human-simplified versions (C2, M = 0.09). Similarly, Figure 9 reveals that the original articles (C1) had the highest kango usage rate (M = 0.31), followed by the human-simplified versions (C2, M = 0.23), and the AI-simplified versions (C3, M = 0.21).

The results demonstrate that both AI and human simplification processes tended to reduce the use of complex vocabulary, including kango, which are often challenging for non-native speakers. However, the AI-simplified versions retained a higher rate of common nouns and a slightly higher rate of intermediate-level vocabulary compared to human-simplified versions. This difference may indicate a nuanced distinction in the approaches to content preservation. It is possible that human simplifiers, guided by their linguistic expertise and pedagogical considerations, prioritized replacing less frequent or more complex words with simpler alternatives, even if it led to a reduction in the overall lexical richness of the text. On the other hand, the AI, trained on vast amounts of text data, appeared more inclined to preserve the original lexical diversity to a greater extent, while still ensuring readability. This lexical analysis reveals a clear contrast with our sentence-level data: while at the sentence level AI created more numerous but shorter sentences, at the lexical level both approaches reduced complexity, though AI preserved more of the original vocabulary diversity and potentially more information content.

To provide a more comprehensive understanding of the AI’s simplification capabilities, we conducted a qualitative review of all 420 AI-simplified texts, focusing on content preservation and accuracy. Our manual examination revealed that the AI successfully maintained the essential information from the original articles in all 420 cases, accurately preserving the core 5W1H elements (who, what, when, where, why, and how). This high level of content preservation is particularly noteworthy given the significant reduction in sentence length and complexity achieved by the AI, as demonstrated in our quantitative analysis. Furthermore, the AI-simplified texts consistently exhibited the ability to replace specialized terms with more accessible explanations, often breaking down complex concepts into simpler components, thereby enhancing the overall clarity and comprehensibility of the texts.^[6]

While the AI demonstrated strong performance in most cases, it is important to acknowledge certain limitations. Out of the initial 424 articles processed, ChatGPT was unable to successfully simplify 4 articles. In these instances, we observed that the AI either failed to complete the simplification task or appeared to misinterpret it as an extreme summarization task, resulting in incomplete or inadequate output. We could not conclusively determine specific reasons for these failures, as they did not follow any clear pattern related to content type or linguistic complexity. This suggests that even advanced AI models like GPT-4 occasionally encounter limitations in text simplification tasks, potentially reflecting broader challenges inherent in large language models regarding task adherence and output consistency, though the occurrence rate in our dataset was relatively low (less than 1 % of the corpus).

In summary, our analyses revealed clear differences in readability scores, sentence-level features, and lexical characteristics across the three corpora. The AI-simplified texts consistently achieved the highest readability scores, followed by human-simplified texts, with original news articles having the lowest scores. At the sentence level, AI broke down content into more numerous but shorter sentences, while human simplifications contained fewer sentences overall and of intermediate length. Lexical analysis showed that both simplification approaches reduced complex vocabulary usage, though AI preserved a higher count of common nouns and higher rates of intermediate-level vocabulary compared to human simplifications. These contrasting patterns suggest fundamentally different approaches to the simplification task, with AI employing systematic transformation while humans appear to rely on synthetic abstraction – distinctions we explore in the next section.^[7]

5 Discussion

In this section, we offer an interpretive analysis of the quantitative findings presented in Section 4, focusing particularly on the contrasting simplification approaches observed between AI and humans. While our discussion is grounded in empirical data, certain aspects – particularly our inferences about the differences between human and AI processing – necessarily involve interpretive elements. Our analysis draws on theoretical frameworks from cognitive linguistics and computational language processing models, integrating these paradigms with our empirical findings to propose explanations for the observed differences in simplification approaches.

Our findings provide strong evidence relevant to all three of our research questions, offering valuable insights into the capabilities of large language models like GPT-4 and the nuances of human cognitive processes in text simplification. As highlighted in our analysis in Section 4, GPT-4 consistently achieved higher readability scores than human experts while preserving a greater degree of the original content (as indicated by morpheme counts) – directly answering our first research question regarding AI performance in Japanese text simplification tasks.

The AI’s sentence decomposition strategy, identified in our quantitative analysis, represents a distinctive approach to simplification that differs fundamentally from human methods. This approach allows the AI to reduce the cognitive load for readers by presenting information in shorter, more digestible units while preserving a greater degree of the original content’s breadth and depth. This strategy demonstrates how AI can effectively balance simplicity and informativeness in ways that human simplifiers – who, as our data suggests (Figure 5), tend to sacrifice content completeness for brevity – generally do not achieve. The effectiveness of this approach directly addresses our second research question regarding the specific simplification strategies employed by AI and humans.

The observed differences in simplification strategies between AI and human experts likely stem from fundamental distinctions in their underlying processing mechanisms and capabilities. Human language processing involves sophisticated abstraction and synthesis abilities – cognitive operations that allow us to distill key points from broader content and prioritize essential information. However, as research in cognitive linguistics suggests (e.g., Chafe 1994; Hasebe 2021), there are also limitations in human working memory and attentional focus during text processing. These cognitive constraints can introduce a significant processing load, particularly when simultaneously attempting to simplify and preserve content. With a limited working memory capacity, estimated to hold only around four chunks of information at a time (e.g., Cowan 2000), humans must often make trade-offs. This frequently leads to prioritizing core message transmission through content reduction rather than attempting the cognitively demanding task of maintaining comprehensive content while simplifying each element.

In contrast, AI models like GPT-4, leveraging transformer architectures with self-attention mechanisms (Vaswani et al. 2017), process information largely in parallel across extensive input sequences. This parallel processing capability allows the AI to implement consistent simplification operations throughout the text while maintaining global coherence and a high degree of content preservation, as observed in our results. These findings on the contrasting nature of human and AI processing – addressing our third research question regarding the implications for understanding human cognitive processes – suggest that the difference lies not merely in scale or efficiency but fundamentally in approach: humans excel at conceptual synthesis and abstraction, often constrained by cognitive limits, while the present AI excels at conducting consistent, systematic transformations across large spans of text.

These fundamental differences in processing approaches become particularly evident when examining practical situations involving text processing and rewriting tasks. For instance, in language education contexts, while skilled language teachers can readily manipulate individual sentences to adhere to simplification guidelines – leveraging their extensive knowledge of lexical and sentence level grammar – they often encounter greater difficulty in managing the complexities of text-level processing. At the sentence level, operations such as converting passive voice to active (“The meeting was canceled by the president” → “The president canceled the meeting”), replacing complex vocabulary with simpler alternatives (“purchase” → “buy”), or breaking down complex sentences with relative clauses into multiple simple sentences can often be performed locally, without extensive consideration of the broader text. In contrast, successful text-level processing requires maintaining coherence across paragraph boundaries, managing anaphoric references consistently, and ensuring logical progression – operations that demand a holistic consideration of the text rather than focusing on isolated sentences. This challenge may be attributed, in part, to the relative lack of a highly formalized and easily applicable framework for “text-level” grammar, compared to sentence grammar – despite foundational work on text structure and coherence (cf. Langacker 2008; van Dijk 2008).

To illustrate these differences more concretely, the following examples from our research data demonstrate this distinction in practice. Example (1) shows an original news article about Facebook’s fake news measures, while examples (2) and (3) represent the human-simplified and AI-simplified versions, respectively. Both human and AI simplifiers effectively replace complex vocabulary with simpler terms – a lexical/sentence-level simplification skill observed in both approaches. However, their approaches to overall text organization differ significantly. As seen in example (2), the human simplifier condenses the entire article into just eight sentences, focusing on core information about Facebook’s new measures against fake news while omitting many secondary details. In contrast, in example (3), the AI simplifier preserves more details from the original article across eleven sentences, largely maintaining the overall structure while simplifying each sentence.

1. Original Text

   

   English Translation

   The American company Facebook, which operates the world’s largest social networking site, has announced measures to prevent the spread of fake news shared among users, including marking suspicious news with a warning. Facebook, which faced criticism that fake news about candidates intentionally spread among users during the recent presidential election may have influenced the election results, announced countermeasures on the 15th. Specifically, external expert organizations will check whether the news is genuine, and if it is deemed suspicious, a warning mark will be attached. Then, when trying to share that news, a screen will appear asking if it’s really okay to share. It will also include a function that allows users to easily report when they feel that the news they are reading is fake. Facebook says these measures will be effective in preventing the spread of fake news, and will start in the United States first, then introduce them in other countries later. In the United States, a survey shows that 44 % of adults check news on Facebook, and there are increasing calls for the same level of responsibility as the media.

2. Human-Simplified Text

   

   English Translation

   This year, there was an election in America to choose the next president. Before the vote, false news about the presidential candidates spread among people using Facebook on the internet. It is said that some people may have decided who to vote for after seeing the false news. Because of this, on the 15th, the Facebook company announced what they will do to prevent false news from spreading. The company will have experts check if the news is true, and will put a mark on news that might not be true. When someone tries to tell their friends about news with this mark, a screen will ask if they really want to share it. The company says they will start doing this in America first, and then in other countries.

3. AI-Simplified Text

   

   English Translation

   Facebook, an American company, has made new rules to stop fake news from being shared. They will put special marks on news that might not be true. This is because in the last presidential election, fake news about candidates spread widely, and this might have affected how people voted. The new rules work like this: experts will check if news is real. If they think it might not be true, they will put a special mark on it. If someone tries to share news with this mark, a screen will ask them if they really want to share it. Also, if people reading news think it’s fake, they can easily report it. Facebook thinks these new rules will help stop fake news from spreading. They will start using these rules in America first, and then in other countries later. In America, 44 % of adults get their news from Facebook. So, many people are saying Facebook should be as responsible as regular news companies.

However, we should refrain from hastily concluding that human-simplified texts are generally lower in quality than AI-simplified ones. The human approach can represent a different kind of text-level processing strategy – one that prioritizes reorganization based on perceived importance rather than comprehensive content preservation. This is evident in how the human simplifier establishes a clearer chronological sequence (presenting the election first, then the problem, then Facebook’s response) while carefully managing topic coherence across a reduced text. This fundamentally different approach demonstrates how human simplification involves conceptual synthesis that deliberately prioritizes key information while omitting secondary details.

The contrasting simplification strategies employed by humans and AI suggest a potentially strong complementary relationship between their respective capabilities in text simplification. While humans often excel at generating original, contextually appropriate text, demonstrating creativity and nuance in their language use, the task of simplifying existing complex text presents unique cognitive challenges, as discussed earlier. These challenges – particularly the need to simultaneously consider multiple textual elements and maintain global coherence while making local simplification decisions – impose a significant cognitive load that appears less burdensome for the parallel processing capabilities of AI. This suggests that AI could effectively augment human abilities in simplification tasks.

In summary, our analysis has highlighted clear differences in how AI and human approaches balance information preservation with simplification: AI tends toward systematic transformation, while humans prioritize conceptual synthesis. These findings not only underscore the distinct strategies involved in text simplification but also offer valuable insights into the interplay between language processing capabilities and cognitive constraints.

6 Conclusions

This study has provided a comparative analysis of AI and human approaches to text simplification in Japanese, directly addressing our three research questions. First, our findings clearly demonstrate that GPT-4 outperforms human experts in Japanese text simplification tasks, achieving significantly higher readability scores (3.69 vs. 2.93) while preserving substantially more original content, as evidenced by higher morpheme counts (332 vs. 205). Second, regarding simplification strategies, we identified fundamental differences: AI tends to decompose complex sentences into multiple shorter units (mean sentence length 21.15 vs. 27.75 for humans), while humans prioritize content reduction and conceptual synthesis, resulting in fewer overall sentences (7.51 vs. 16.10 for AI) but greater content transformation.

These contrasting approaches also illuminate our third research question regarding the implications for understanding human cognitive processes in language. The human approach to simplification – characterized by significant content reduction and conceptual synthesis – likely reflects inherent limitations in working memory and attentional focus when undertaking the complex simplification task. By contrast, AI’s parallel processing capabilities allow it to implement consistent simplification operations across larger spans of text while maintaining global coherence and preserving more information, tasks that pose cognitive challenges for humans.

The fundamental differences observed in our study suggest a potentially powerful complementary relationship between human and AI capabilities in text processing. While humans often excel at generating contextually appropriate text with creativity and nuance, the specific task of simplifying existing complex content presents unique cognitive challenges. These challenges – particularly maintaining global coherence while making local simplification decisions – appear better suited to AI’s processing architecture. This complementarity suggests promising applications in Japanese language education and accessibility initiatives. For instance, AI could handle initial, systematic simplification while preserving content breadth, leaving human educators to focus on adding culturally appropriate explanations and essential pedagogical refinements that require deeper contextual understanding and pedagogical expertise.

Future research could explore optimal frameworks for synergistic human-AI collaboration in text simplification, extending this investigation to other languages to understand how simplification strategies might vary across diverse linguistic structures. Additionally, the cognitive insights gained from comparing AI and human simplification strategies could directly inform the development of more effective language teaching methodologies and accessibility tools. By harnessing the distinct strengths of both human and artificial intelligence, we can work toward creating more accessible content for language learners, foreign residents, and others who require simplified Japanese materials, ultimately promoting greater inclusivity in communication.