2.1. Low-level word segmentation cues.

Low-level word segmentation cues are perceptually salient, visually-based markers that provide explicit spatial demarcation of word boundaries in text [19]. These cues operate primarily at the level of orthographic processing, preceding lexical access, and function independently of linguistic context or semantic information [20,21]. In alphabetic scripts, the inter-word space is a typical low-level visual word segmentation cue. It explicitly marks the boundaries between words, helping readers to perform word segmentation quickly [15].

Low-level cues primarily assist word segmentation through directly visible visual boundary markers in the text, using physical separation. Research across multiple writing systems has demonstrated the effectiveness of various low-level segmentation markers, including inter-word spacing in alphabetic languages [3,22], visual highlighting and color alternation in unspaced scripts [23,24], and other spatial modifications [25,26]. They manifest in the following forms: 1. Inter-word space markers: This is the most direct form of visual separation. When spaces are inserted into Chinese text to mark word boundaries (e.g., “中文 阅读” [Chinese reading]), research has found that inter-word spaces significantly enhance the reading efficiency of beginners, such as primary school students. The mechanism is that they provide clear visual boundaries for lexical recognition [5,15]. Experimental evidence showed that Chinese third-grade students found reading Chinese text with inter-word spaces to be of comparable difficulty to reading normal, unspaced text. However, non-word spaces (e.g., “中 文阅 读” [C-hinese rea-ding]) severely interfered with students who had lower reading skills [15]. 2. Alternating color markers: Color-based boundary demarcation (e.g., alternating red and green) avoids the changes in text spatial distribution caused by spaces. Studies found that word-appropriate color marking maintains reading efficiency comparable to normal text; however, non-word color marking interfered with reading [27,28]. Critically, this color marking effect only affects the visual information processing stage and is not influenced by word frequency, which suggests its role is confined to the early stage of lexical recognition [27]. This pattern has been replicated in Tibetan reading, where color alternation markings facilitated lexical recognition but did not affect saccade target selection [28]. 3. Gray bar markers: Using gray bars to replace spaces for marking boundaries maintains the spatial distribution of the sentence. Experiments showed that their effect is similar to that of inter-word spaces but avoids the interference on reading caused by changes in physical space [15].

2.2. High-level word segmentation cues.

Unlike low-level word segmentation cues such as inter-word spaces, high-level word segmentation cues refer to the grammatical and semantic information that readers rely on to assist in word segmentation during the reading process. According to theoretical frameworks of word segmentation in unspaced languages [17,29], high-level cues operate through interactive activation processes involving both bottom-up character-level processing and top-down lexical and contextual information [30]. Li et al.‘s (2009) computational model of Chinese word segmentation proposes that word identification and segmentation occur simultaneously through competition among candidate words activated by character-level input, modulated by lexical frequency and contextual predictability [1]. This interactive process contrasts with the serial word identification assumed in models of alphabetic reading like E-Z Reader [31,32].

For instance, when we see the string of unspaced English characters—Thesunriseseverymorning—we can still understand the meaning of the sentence, albeit with some difficulty: The sun rises every morning. This is because we search for familiar character combinations within this string and attempt to segment it into meaningful and plausible words to comprehend its meaning. In this process, we are using high-level word segmentation cues. These cues can help readers correctly divide words in a sentence by leveraging semantic understanding, grammatical structure, and word familiarity, especially when faced with ambiguity or a lack of clear word boundaries [33]. Empirical evidence from Chinese reading demonstrates that readers utilize statistical regularities, morpheme positional probabilities, and lexical context to segment words during natural reading [18,34]. Similarly, research in Tibetan reading suggests that, despite the absence of inter-word spacing, readers rely on linguistic cues to achieve efficient word segmentation [4,35]. For languages like Tibetan and Chinese, which naturally lack low-level word segmentation cues, high-level cues are particularly important.

Among high-level word segmentation cues, morpheme positional frequency is the most extensively studied. When we read a stretch of unspaced English text, we are highly likely to treat “pre-” as the beginning of a word and “-ing” as the end of a word. This is because “pre-” commonly appears at the beginning of words, while “-ing” commonly appears at the end. This process of judging word boundaries based on whether a morpheme more frequently appears at the beginning or end of a word is word segmentation through morpheme positional frequency. In languages with a finite number of morphemes, such as Chinese and Tibetan, morpheme positional frequency is defined as the statistical frequency of a Chinese character (or a Tibetan morpheme) appearing in a specific position (initial or final) within a two-character word (or two-morpheme word). In Chinese, the calculation method for morpheme positional frequency is: the number of two-character words where a specific character appears in the initial position, divided by the total number of two-character words that the character helps to form [11]. For example, the character “精” (jīng, essence) predominantly appears at the beginning of two-character words, yielding a high initial positional frequency (>0.85), as in words like “精神” (jīngshén, spirit) [11].

Characters with high positional frequency (like “精”, which often appears in the initial position) provide readers with implicit word boundary cues, helping them to segment words quickly [11,36]. Experimental studies have shown that when the initial and final morpheme positional frequency of a novel word are consistent with its actual position (e.g., the initial morpheme often appears at the beginning of words, and the final morpheme often appears at the end), lexical recognition is fastest [5,37]. Conversely, inconsistent positional frequency (e.g., an initial morpheme that often appears at the end of words) significantly prolongs fixation duration [37]. This finding revealed the cognitive mechanism of morpheme positional frequency as an internal statistical cue: through long-term language exposure, readers internalize the positional distribution patterns of characters within words and automatically activate this statistical knowledge during lexical recognition, thereby facilitating the rapid identification of word boundaries.

3.1. The role of low-level word segmentation cues in word segmentation.

3.1.1. The Role of Spaces in Word Segmentation: When reading alphabetic scripts like English and Spanish, readers heavily rely on the inter-word space as a low-level word segmentation cue [38,39]. This reliance on spacing is observed broadly across alphabetic writing systems that employ inter-word spaces, including not only Latin-based languages (English, Spanish, German) but also Greek, Cyrillic-based languages (Russian), and Hebrew [40,41]. In these languages with inter-word spaces, spaces have three core functions: providing word boundary cues [39], which reduces lexical ambiguity and facilitates parallel word identification in parafoveal vision [42]; guiding eye movement landing positions [38], enabling readers to target optimal viewing positions within words and plan efficient saccades [43]; and reducing letter visual crowding [44], thereby improving letter recognition accuracy through lateral interference reduction [45].

When spaces are removed, the reading speed in both languages decreases significantly, by 40−70% [46], and this leads to disordered eye movement patterns [38]. Sheridan et al. (2016) conducted a study on English, comparing eye movement trajectories for normal spaced text and unspaced text. The study recruited native English-speaking university students (n = 24) as participants and they found that removing spaces led to: a decrease in reading speed, as the amount of information processed per unit of time was reduced (average decrease of 20%−30%) [47]; prolonged fixation durations, with single fixation durations increasing by 15%−25%, reflecting increased difficulty in lexical recognition [47]; abnormal skipping patterns, with the word skipping rate decreasing by 30%, indicating that parafoveal processing was hindered [47]; and shortened saccade distances by 10%−15%, with the initial fixation position being closer to the beginning of the word (a shift of 0.3-0.5 letter positions) [47]. In unspaced text, the word frequency effect appeared later (emerging at 60 ms in normal text, but delayed to 100 ms in unspaced text), which showed that lexical recognition speed was slower [47].

Similar to the findings of Sheridan et al., an experiment by Juhasz et al. (2008) found that in English reading, the absence of spaces increased the first fixation duration on short words (≤4 letters) by 22% [39]. This particularly strong effect for short words occurs because shorter words have fewer internal orthographic features to distinguish them from adjacent words, making them more dependent on external boundary cues (spaces) for rapid identification [48]. Additionally, short words typically receive less parafoveal preview benefit, making immediate visual segmentation cues more critical [49]. This further demonstrated the primary role of spaces for word segmentation in English reading.

Research in Spanish yielded similar results to those in English, demonstrating broader generalizability. Toshima et al. (2011) used Spanish as the research language, asking native Spanish-speaking participants to read unspaced Spanish sentences. The results from multiple-choice comprehension questions administered after each sentence showed: a 35% increase in word segmentation errors and a decrease in comprehension accuracy, with the correctness rate for reading comprehension questions on unspaced text dropping by 18% [50]. The absence of spaces forced readers to allocate more cognitive resources to word segmentation, which crowded out resources for semantic integration [50]. This indicates that in languages with inter-word spaces, the space is a critical word segmentation cue that is vital for reading comprehension.

3.1.2. The Role of the Saliency of Kanji Characters in Word Segmentation: For scripts that do not have inter-word spaces, other low-level word segmentation cues may exist. For instance, in Japanese text, the visual contrast between Kanji (ideographic characters) and Kana (syllabic scripts) serves as a salient boundary marker [51,52]. Early research by Kajii, Nazir, and Osaka (2001) demonstrated that Kanji characters, due to their greater visual complexity and morphographic nature, function as visual anchors that guide eye movements and facilitate word boundary identification in mixed-script Japanese text [52]. The boundaries between Hiragana, Katakana, and Kanji can serve as a basis for judging word boundaries.

Fujii and Ishikawa (2001) found that the word segmentation accuracy for combinations like “CK” (Kanji + Katakana) and “CA” (Kanji + Alphabet) was higher, reaching 95% in a Japanese-English cross-language information retrieval task where native Japanese speakers (n = 30) segmented compound technical terms [53]. For instance, in the character string “東京タワー” (Tokyo Tower), it is easier for us to identify the word boundary between “東京” and “タワー” [53]. This is because Kanji characters have a higher visual complexity compared to Hiragana and Katakana [54], making them easy to form “visual anchors” in the text. An eye movement experiment [14] showed that native Japanese speakers did not need space assistance in mixed Kanji-Kana text because Kanji provides word boundary cues; however, inserting spaces into purely Kana text could improve reading efficiency by 26% (measured as a 26% reduction in total sentence reading time). This indirectly suggests that Japanese readers rely on the saliency of Kanji for word segmentation.

3.2. High-level word segmentation cues: The role of morpheme positional frequency in word segmentation.

Morpheme positional frequency is a high-level word segmentation cue (e.g., morpheme positional frequency), often found in languages without low-level cues (e.g., spaces between words), such as Chinese, Tibetan, Thai, and the Yi script. Research on the effect of morpheme positional frequency on word segmentation has so far been concentrated on Chinese.

3.2.1. The Effect of Morpheme Positional Frequency in Natural Reading: First, morpheme positional frequency (the statistical frequency of a character appearing in a specific position) affects word segmentation during natural reading. Liang Feifei et al. conducted two experiments with 48 adults and 48 third-grade children who were native speakers of Chinese as participants. In Experiment 1, they selected two target words with similar meanings to control for semantic processing difficulty and contextual predictability, ensuring that any observed differences in reading times would be attributable to positional frequency manipulation rather than semantic factors [55]. They then manipulated the frequency of the initial character’s appearance. For example, in “湖水” (húshuǐ, lake water) and “泉水” (quánshuǐ, spring water), the character “湖” (hú) has a higher frequency of appearing at the beginning of a word, while “泉” (quán) has a lower frequency. The target word was then embedded in the same sentence, and the participants’ eye movements were observed as they read the two sets of sentences. The results showed that the initial morpheme positional frequency did not have a significant effect on word segmentation or lexical recognition. In Experiment 2, they manipulated the second character of the target word. The results revealed that the final morpheme positional frequency of the target word had an effect on the early stage of lexical recognition in adult reading, meaning it could help readers successfully complete word segmentation and lexical recognition. In children’s reading, however, the effect of this frequency appeared in the later stage. This may reflect differences in reading proficiency and cognitive processing between adults and children. Experiment 1 and Experiment 2 confirmed that the positional frequency of the final morpheme (but not the initial morpheme) has a significant impact on word segmentation and lexical recognition [55].

3.2.2. The Effect of Morpheme Positional Frequency in the Acquisition of Novel Words: Second, morpheme positional frequency affects word segmentation during the acquisition of novel words. Research in this area has commonly used the novel word learning paradigm. In this paradigm, novel words are created according to certain rules, and the positional frequency of their constituent parts (i.e., characters) is manipulated, including the initial morpheme positional frequency and final morpheme positional frequency of the novel word. The newly created novel words are then embedded in sentences, and participants are asked to read and learn them. Eye movement measures—including fixation duration, gaze duration, and total reading time—are considered reliable indicators of word segmentation processes because they provide millisecond-by-millisecond temporal indices of ongoing cognitive processing during natural reading [56,57]. Specifically, longer fixation durations and increased refixations typically reflect difficulty in word segmentation and lexical identification, while shorter fixations indicate more efficient boundary detection and lexical access [40,58]. These measures have been extensively validated as sensitive indices of lexical processing difficulty and word boundary ambiguity [59,60]. By observing participants’ eye movements during this process, one can infer the potential influence of morpheme positional frequency on word segmentation in the acquisition of novel words [61].

Liang Feifei conducted a study with 15 university students who were native Chinese speakers and constructed three types of novel words based on the positional frequency of each character. In the first type of novel word, the position of the constituent characters was consistent with their positional frequency; in the second type, the position was inconsistent; in the third type, the frequency of the constituent characters appearing in the initial position was similar to their frequency of appearing in the final position. Each novel word was embedded in a sentence. The experiment included two phases: a learning phase and a testing phase. In the learning phase, participants were only required to read the sentences containing the novel words. In the testing phase, participants were not only required to read the sentences but also to complete a semantic categorization task, such as determining whether the novel word belonged to the category of animals, plants, furniture, or another given category. The learning outcomes of the participants could be evaluated through this task. During the learning and testing phases, participants’ eye movement behaviors were recorded using an eye-tracker. Based on an analysis of skipping rate, fixation duration, fixation count, and fixation position, Liang Feifei concluded that in Chinese reading, morpheme positional frequency is an effective linguistic word segmentation marker. Its effect spans the entire process of lexical recognition, with a broad scope and long duration of action [10].

3.2.3. The Effect of Morpheme Positional Frequency in Overlapping Ambiguous Strings: Furthermore, morpheme positional frequency affects word segmentation in overlapping ambiguous strings. Cao Haibo et al. manipulated the initial and final morpheme positional frequency of overlapping ambiguous strings (e.g., “邮差距” yóu-chā-jù, which can be segmented as “邮差” yóu-chāi, postman, or “差距” chā-jù, gap) and asked participants to quickly report the pronunciation of the middle character. The results showed that readers tended to report the pronunciation of the character on the side with the higher morpheme positional frequency. In another of their experiments, they manipulated the initial and final morpheme positional frequency of overlapping ambiguous strings (e.g., “惹祸害” rě-huò-hài, which can be segmented as “惹祸” rě-huò, to cause trouble, or “祸害” huò-hài, to harm) and asked participants to report one word from the ambiguous string. The results found that readers tended to report the word on the side with the higher morpheme positional frequency. The results of both experiments indicate that morpheme positional frequency is an important linguistic word segmentation cue for Chinese readers. Readers can use the initial and final morpheme positional frequency to segment overlapping ambiguous strings, and both cues play a role in the process of lexical recognition [62].

3.2.4. The Effect of Morpheme Positional Frequency in Reading Aloud: Finally, morpheme positional frequency also influences word segmentation during reading aloud. Lian Kunyu et al., through two experiments, manipulated the level (high or low) of initial and final morpheme positional frequency to investigate whether Chinese readers use this information for word segmentation during reading aloud. The results found that under reading-aloud conditions, participants could utilize initial morpheme positional frequency to perform word segmentation [63].

5.1. Directly examining the initial stage of word segmentation and lexical acquisition.

The acquisition of novel words involves integrating entirely new vocabulary into a learner’s lexicon, a process highly dependent on word segmentation skills. Empirical studies in Chinese have demonstrated that morpheme positional frequency significantly affects novel word learning during natural reading [55,61]. Morpheme positional frequency, as an implicit cue, can help learners determine word boundaries, and the novel word learning paradigm provides a direct opportunity to observe the effect of this cue. In languages without explicit word boundaries, learners may rely on high-positional-frequency morphemes to infer the start and end positions of words, a process that is particularly evident in the acquisition of novel words [55,61].

5.2. Controlling experimental variables to isolate the effect of morpheme positional frequency.

In natural reading tasks, the effect of morpheme positional frequency is often confounded with other factors (such as word frequency and contextual predictability), making it difficult to analyze its role in isolation. For example, in Chinese, the morpheme 湖 (hú, ‘lake’) has a high initial positional frequency, while the morpheme 泉 (quán, ‘spring’) has a low one. A simple comparison between existing words like 湖水 (húshuǐ, ‘lake water’) and 泉水 (quánshuǐ, ‘spring water’) is confounded by factors such as their differing word frequencies and orthographic complexity (i.e., stroke counts). To isolate the effect of positional frequency, a researcher could create two novel words, such as by pairing each morpheme with the same neutral character. By ensuring these two novel words are presented an equal number of times, any observed difference in processing can be more reliably attributed to the positional frequency of the initial morphemes (湖 vs. 泉). This controlled environment can effectively isolate the independent effect of morpheme positional frequency, providing clearer causal evidence [55,61].

5.3. Real-time capture of cognitive processing.

The novel word learning paradigm, when combined with eye-tracking technology, can capture the cognitive dynamics of learners in real-time during word segmentation and lexical recognition. This paradigm often presents a novel word in different sentences six or even more times. By analyzing indicators such as skipping rates and fixation durations during the multiple exposures to the novel word, it is possible to precisely reveal how morpheme positional frequency affects the learning and processing of novel words. Such real-time data are more difficult to obtain in natural reading, as it rarely creates six similar reading processes [55,61].

In summary, using the novel word learning paradigm to study morpheme positional frequency in Tibetan reading will be particularly valuable. Given that Tibetan, like Chinese, lacks inter-word spacing, investigating whether Tibetan readers utilize morpheme positional frequency cues during novel word acquisition will be an efficient and robust method and could reveal fundamental similarities and differences in word segmentation mechanisms across unspaced writing systems.

Participants.

To determine an appropriate sample size, we conducted an a priori power analysis using GPower (Version 3.1.9.2) [70]. The study employed a 2 × 2 within-participants design, so we based our calculation on the “ANOVA: Repeated measures, within factors” model in F-tests. The parameters were set as follows: effect size f = 0.25 (medium effect), α = 0.05, statistical power (1-β) = 0.80, number of groups = 1, number of measurements = 4, and correlation among repeated measures = 0.5, assuming sphericity (nonsphericity correction ε = 1). The calculation showed that the minimum sample size required to achieve the desired power was 24. It should be noted that the subsequent analysis used the lme4 package in R to conduct a 2 × 2 linear mixed model (LMM) analysis of fixed effects [71]. Although this analysis is more suitable for within-participants designs, LMM was chosen over traditional ANOVA as it offers greater flexibility, particularly in its ability to simultaneously account for both participant and item variability through random effects and to handle missing data points without excluding entire participants [71]. Therefore, the sample size calculated by GPower might have a slight margin of error. Consequently, we ultimately recruited 60 participants. After excluding samples (n = 4) with an eye-tracking data collection rate below 85%, the number of valid participants was 56. These 56 native Tibetan speakers (28 male, 28 female, all of whom were also proficient in Mandarin Chinese) were from Tibetan communities in Shanxi Province, China, with an age range of 18–55 years (M = 34.7, SD = 9.1). All participants were right-handed (a criterion to minimize variability in brain lateralization for language processing), had normal or corrected-to-normal vision, and had no history of neurological disorders. All participants signed an informed consent form before the experiment and received 50 RMB as compensation after the experiment.

Ethics statement.

This prospective study was approved by the university-level Science and Technology Ethics Committee, Shanxi Normal University (Approval No. 2025−0701, Date: July 20, 2025). Participant recruitment started on 20 July 2025 and ended on 30 July 2025, and all study procedures were conducted at Shanxi Normal University. All participants were adults and provided written informed consent prior to participation; no minors were enrolled. The study complied with institutional guidelines and the Declaration of Helsinki. All data were anonymized at the point of collection and stored on encrypted institutional servers. De-identified data and materials are publicly available at DOI: https://doi.org/10.6084/m9.figshare.30314464.v3.

Experimental design.

This experiment used a 2 × 2 within-participants design. The two factors were Factor A (initial morpheme positional frequency) with two levels (A1: high, A2: low) and Factor B (reading phase) with two levels (B1: learning phase, B2: testing phase). The 2 × 2 within-participants design directly tests both main effects and their interaction with high efficiency and statistical power, and it follows prior Chinese-language studies employing the same paradigm [8,10,37,55,61,72–74]. Each item (trial) comprised six sentences: the first three constituted the learning phase and the subsequent three constituted the testing phase [73,74]. The order of these six sentences was fixed (no within-trial randomization or counterbalancing). To control order effects at the session level, the order of trials (i.e., the six-sentence item blocks) was randomized independently for each participant [73,74].

Experimental materials.

This study referenced the modern Tibetan corpus created by Robert Barnett et al. [75]. This corpus was extracted from Tibetan newspapers published in recent years and has been processed with word segmentation. We conducted a statistical analysis of all words in this corpus and selected 11 morphemes with high initial morpheme positional frequency and low final morpheme positional frequency (the probability of these morphemes appearing at the beginning of a word exceeds 85%, and the probability of appearing at the end is less than 15%); these were designated as Morpheme 1, following conventions adopted in prior Chinese-language studies on morpheme positional frequency [73,74]. We also selected 10 morphemes with low initial morpheme positional frequency and high final morpheme positional frequency (the probability of these morphemes appearing at the beginning of a word is less than 15%, and the probability of appearing at the end exceeds 85%); these were designated as Morpheme 3. Simultaneously, we selected 15 morphemes with comparable initial and final morpheme positional frequencies (the probability of these morphemes appearing at the beginning and end of a word was between 45% and 55%); these were designated as Morpheme 2. We also controlled for the frequency of these morphemes in Tibetan texts and matched the length of the novel words by character count. Finally, we combined four Morpheme 1s with four Morpheme 2s to create four Tibetan novel words, and combined four Morpheme 3s with four Morpheme 2s to create another four Tibetan novel words. Both sets of novel words used the same Morpheme 2s. The order within each novel word was fixed as M1–M2 (placing M1 word-initial) and M2–M3 (placing M3 word-final), respectively. The same set of Morpheme 2s appeared word-final in the M1–M2 items and word-initial in the M2–M3 items. These novel words were evaluated by 10 native Tibetan speakers (who did not participate in the subsequent experiment) to confirm that they had no actual meaning. They performed a binary real-word judgment, and there was 100% consensus that all items had no attested meaning. Subsequently, each novel word was embedded in six highly constraining sentences. Sentence constraint was established by expert judgment, consistent with conventions in related Chinese-language studies [73,74]. In highly constraining sentences, participants could infer the meaning of the novel word from the sentence context. This process generated eight sets of sentences, with each set containing six sentences. We present an example (one set of sentences) in Fig 1. The procedure for creating our experimental materials was consistent with previous studies [73,74].

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. These sentences can be translated to: There is a zopax at his front door.

The lazy zopax is sleeping over there. Have you seen cartoons about zopax and mice? The smart zopax caught a mouse. He bought a new toy for the zopax. She took the zopax to the pet hospital. Through these 6 sentences, we can learn that zopax should refer to a type of pet animal, and most likely a type of cat. Figures 1 are original. English glosses are provided for reader clarity.

https://doi.org/10.1371/journal.pone.0337615.g001

Each set of sentences was paired with a corresponding multiple-choice question to test whether the participant truly understood the novel word. An example of the multiple-choice question can be seen in Fig 2.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. On this image, there is a multiple-choice question.

The question part is: The word “ zopax” most likely belongs to which of the following categories? The four options are: A. student, B. insect, C. cat, D. book. Based on reading and inferring from the previous 6 sentences, we can determine that the correct answer is C. Fig 2 are original. English glosses are provided for reader clarity.

https://doi.org/10.1371/journal.pone.0337615.g002

Apparatus.

We used a Tobii X3-120 eye-tracker, which has a sampling frequency of 120 Hz [76]. The eye-tracker was equipped with Tobii Pro Lab, a professional software for designing eye-tracking experiments [76]. We used Tobii Pro Lab to design and conduct the experiment. Stimuli were presented on a 15-inch DELL OLED display with a resolution of 1920 × 1080 pixels. The monitor was mounted on an adjustable stand. A height-adjustable headrest was used to stabilize the participant’s head at a distance of 60 cm from the screen. The laboratory was soundproofed and light-isolated.

Procedure.

Participants were tested individually. First, a five-point calibration was performed, with an average error of less than 0.3°. After successful calibration, instructions were presented. Participants proceeded to the practice trials after confirming they understood the task. The practice phase comprised 7 screens (6 sentences + 1 multiple-choice question). The formal experiment then began. Per participant, the formal phase included 8 sets × 6 sentences (48 reading screens) plus 8 multiple-choice questions (total 56 screens). One sentence was presented per screen. After reading, participants pressed the spacebar to advance to the next page. Participants read the sentences set by set. Within each set, the order of sentences was fixed. However, the presentation order of the sets was randomized. Following previous studies, we defined the reading of the first three sentences in a set as the learning phase and the reading of the last three sentences as the testing phase [8,61]. After reading a set of sentences, participants were presented with a multiple-choice question. This question required them to determine the semantic category of the learned novel word based on the descriptions in the six sentences. Participants answered the multiple-choice questions on the screen using the Right Ctrl, Left Arrow, Down Arrow, and Right Arrow keys to represent A, B, C, and D, respectively. The Right Ctrl, Left Arrow, Down Arrow, and Right Arrow keys are four consecutive keys on our keyboard; this layout was chosen simply for the convenience of the participants and had no special significance. Only accuracy was analyzed; response times for these multiple-choice questions were not modeled. After answering, participants proceeded to the next set of sentences. The experiment concluded after all sentences were read. A flowchart of the experimental procedure is shown in Fig 3. No additional attention checks were administered beyond the built-in comprehension questions. The entire experiment lasted approximately 8 minutes on average.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. The flowchart of the experimental procedure.

https://doi.org/10.1371/journal.pone.0337615.g003

Measures.

Based on previous research on the acquisition of novel words, we identified three categories of sensitive measures in this field. The first category includes measures reflecting saccade target selection, such as skipping rate and initial fixation position [56,77]. Skipping rate represents the percentage of trials in which a participant did not fixate on the target word during a single sentence presentation [56,77]. Initial fixation position refers to the horizontal position of the first fixation landing within the target word [56,77]. To eliminate the influence of Tibetan word length differences, we converted the horizontal distance of this landing point from the word’s left boundary into a percentage of the word’s width (0% = beginning of the word, 100% = end of the word). These two measures can reflect the participant’s detection and acquisition of morpheme positional frequency in the parafoveal region [56].

The second category includes measures reflecting the early stage of lexical recognition, such as first fixation duration and gaze duration [57,77]. First fixation duration is the duration of the first fixation on the target word [56,77]. It is an indicator of the very initial stage of lexical recognition [57,77]. Gaze duration is the sum of fixation durations on the target word from the first fixation until the gaze first moves away from it [56,77]. It reflects the time required for the initial processing of the target word [57,77].

The third category includes measures reflecting the later stage of lexical recognition, such as number of fixations, total fixation time, and regression pass reading time [73,77]. Number of fixations is the count of fixations on the target word, reflecting the overall processing difficulty of the target word [77,78]. Total fixation time is the sum of all fixation durations on the target word. Compared to gaze duration, total fixation time includes fixations made when the participant returns to the target word after briefly looking away. Therefore, it accounts for instances where the participant experiences confusion after reading past the target word and has to return to it [77,79]. Regression pass reading time is the sum of all fixation durations from the first fixation on the target word until the gaze last moves to the right of it. This measure includes not only fixations on the target word but also fixations to the left of it after the first fixation [77,79].

In summary, we used a total of seven measures in our experiment, allowing for a more comprehensive analysis of participants’ eye movement behavior. For all time-based measures, we coded trials without a fixation on the target area (TOI) as missing (rather than 0) and excluded them from the time-based analyses, consistent with the standard definitions of these measures and common practice in Chinese reading research [73,74,77]. For regression pass reading time, we likewise excluded trials in which the target was skipped on the first pass and only fixated after a regressive saccade, consistent with the standard definitions of these measures and common practice in Chinese reading research [73,74,77]. We acknowledge that these exclusions may introduce bias when skipping differs by condition.

Results.

The average answer accuracy was 98.5%, indicating that the participants carefully read and understood the novel words and sentences. Following the existing studies [8,61], data were excluded based on the following three criteria: (1) premature or incorrect key presses that interrupted sentence presentation, (2) invalid data due to loss of tracking, (3) fixation durations of less than 0.08s. Consequently, 0.1% of total eye-movement data points (fixations) in the analysis were excluded due to being invalid. Before analysis, we used the differential evolution algorithm from Python’s SciPy library to calibrate the data of participants by optimizing per-participant calibration to reduce residual drift [80]. To assess robustness, we re-fit all models with versus without the 0.1% flagged fixations; the direction and statistical significance of the fixed effects were unchanged. The raw and processed data needed to reproduce this analysis are available at DOI: https://doi.org/10.6084/m9.figshare.30314464.v3.

Data analysis was conducted in the R environment (version 4.4.3). For continuous dependent variables (first fixation duration, gaze duration, number of fixations, total fixation time, regression pass reading time, and initial fixation position), we fitted Linear Mixed-Effects Models (LMMs) using the lme4 [71] and lmerTest packages. The models included fixed effects for morpheme positional frequency (Factor A: two levels), reading phase (Factor B: two levels), and their interaction. Random intercepts for participants were included as random effects: DV ~ FactorA * FactorB + (1|Recording). P-values and degrees of freedom were estimated using the Satterthwaite approximation implemented in lmerTest. For binary dependent variables (skipping rate), we fitted Generalized Linear Mixed-Effects Models (GLMMs) with a binomial distribution and logit link function, using the same fixed and random effects structure: Binary_DV ~ FactorA * FactorB + (1|Recording). Significance tests for GLMMs were conducted using Type III Wald chi-square tests via the Anova() function from the car package. Models were fitted using the bobyqa optimizer with a maximum of 200,000 function evaluations. Although we report raw means and standard deviations in tables for descriptive purposes, statistical inferences were based on the mixed-effects models described above. We did not apply corrections for multiple comparisons across the seven dependent variables, as each measure was analyzed separately and reflects distinct stages of lexical processing with theoretically motivated predictions.

To make the experimental results clearer, we have compiled the findings of this experiment into a descriptive statistics table as shown in Table 1, a significance test table as shown in Table 2, and a column chart as illustrated in Fig 4.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Means and standard deviations of eye movement indicators in experiment 1 (standard deviations in parentheses).

https://doi.org/10.1371/journal.pone.0337615.t001

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Statistical analysis of eye movement indicators in experiment 1. * p < 0.05; ** p < 0.01; *** p < 0.001.

https://doi.org/10.1371/journal.pone.0337615.t002

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. Normalized mean values of dependent variables across experimental conditions in Experiment 1.

Each panel represents a different dependent variable (abbreviated as shown in panel headers). Bars show mean values normalized to the maximum value within each variable across both experiments (maximum = 1.0), enabling direct comparison between Experiments 1 and 2. Error bars represent ±1 standard deviation after normalization. Four experimental conditions are shown: A1B1, A1B2, A2B1, and A2B2, representing the factorial combination of two factors (A and B) at two levels each.

https://doi.org/10.1371/journal.pone.0337615.g004

Our analysis revealed significant main effects for both initial morpheme positional frequency (factor A) and reading phase (factor B), but no significant interaction between them across all eye-tracking measures.

The main effect of initial morpheme positional frequency (factor A) was significant, primarily influencing later-stage processing and eye movement targeting. Specifically, novel words with a low initial morpheme positional frequency were more difficult to process. This difficulty manifested as a more leftward initial fixation position (p = .002), a greater number of fixations (p = .030), longer total fixation time (p = .017), and longer regression pass reading time (p = .023) compared to words with a high-frequency initial morpheme. Notably, this factor did not significantly affect early processing measures like first fixation duration or gaze duration.

The main effect of reading phase (factor B) was also significant, indicating a clear learning effect. As participants moved from the learning phase to the testing phase, their processing of the novel words became significantly more efficient. This was reflected in a higher skipping rate (p = .025), shorter first fixation duration (p = .015), shorter gaze duration (p < .001), fewer number of fixations (p < .001), shorter total fixation time (p < .001), and shorter regression pass reading time (p < .001). However, this factor did not have a significant effect on the measure of initial fixation position.

Finally, except for the initial fixation position showing a marginally significant interaction effect (p = .092), all other indicators showed no significant interaction effects between initial morpheme positional frequency and reading phase (all p > 0.1), suggesting largely independent contributions of positional frequency and learning phase to eye-movement behavior.

Participants.

As in Experiment 1, 56 valid participants were drawn from the same pool as in Experiment 1: 56 of the 60 individuals recruited for Experiment 1 took part in Experiment 2. Four participants with poor gaze-tracking quality in Experiment 1 (valid sample rate < 85%, approximately 30–40%) were not invited back for Experiment 2; no additional exclusions were made. The order of participation (Experiment 1 first vs. Experiment 2 first) was randomized by lot, with a one-week interval between sessions. Given this counterbalancing and the washout interval, potential repeated-exposure and order effects are expected to be minimized and to balance out across participants. Recruitment procedures, inclusion criteria, and demographics matched those of Experiment 1.. All participants signed an informed consent form before the experiment and received 50 RMB as compensation after the experiment.

Ethics statement.

This prospective study was approved by the university-level Science and Technology Ethics Committee, Shanxi Normal University (Approval No. 2025−0701, Date: July 20, 2025). Participant recruitment started on 20 July 2025 and ended on 30 July 2025, and all study procedures were conducted at Shanxi Normal University. All participants were adults and provided written informed consent prior to participation; no minors were enrolled. The study complied with institutional guidelines and the Declaration of Helsinki. All data were anonymized at the point of collection and stored on encrypted institutional servers. De-identified data and materials are publicly available at DOI: https://doi.org/10.6084/m9.figshare.30314464.v3. (The same approval covered both experiments.)

Experimental Design.

This experiment used a 2 × 2 within-participants design. The two factors were Factor A (final morpheme positional frequency) with two levels (A1: high, A2: low) and Factor B (reading phase) with two levels (B1: learning phase, B2: testing phase). The 2 × 2 within-participants design directly tests both main effects and their interaction with high efficiency and statistical power, and it follows prior Chinese-language studies employing the same paradigm [8,10,37,55,61,72–74]. Each item (trial) comprised six sentences: the first three constituted the learning phase and the subsequent three constituted the testing phase [73,74]. The order of these six sentences was fixed (no within-trial randomization or counterbalancing). To control order effects at the session level, the order of trials (i.e., the six-sentence item blocks) was randomized independently for each participant [73,74].

Experimental Materials.

As in Experiment 1, this study referenced the modern Tibetan corpus created by Robert Barnett et al. [75]. This corpus was extracted from Tibetan newspapers published in recent years and has been processed with word segmentation. We conducted a statistical analysis of all words in this corpus and selected 11 morphemes with high initial morpheme positional frequency and low final morpheme positional frequency (the probability of these morphemes appearing at the beginning of a word exceeds 85%, and the probability of appearing at the end is less than 15%); these were designated as Morpheme 1, following conventions adopted in prior Chinese-language studies on morpheme positional frequency [73,74].. We also selected 10 morphemes with low initial morpheme positional frequency and high final morpheme positional frequency (the probability of these morphemes appearing at the beginning of a word is less than 15%, and the probability of appearing at the end exceeds 85%); these were designated as Morpheme 3. Simultaneously, we selected 15 morphemes with comparable initial and final morpheme positional frequencies (the probability of these morphemes appearing at the beginning and end of a word was between 45% and 55%); these were designated as Morpheme 2. We also controlled for the frequency of these morphemes in Tibetan texts and matched the length of the novel words by character count. Finally, we combined four Morpheme 2s with four Morpheme 3s to create four Tibetan novel words, and combined four Morpheme 2s with four Morpheme 1s to create another four Tibetan novel words. Both sets of novel words used the same Morpheme 2s. The order within each novel word was fixed as M1–M2 (placing M1 word-initial) and M2–M3 (placing M3 word-final), respectively. The same set of Morpheme 2s appeared word-final in the M1–M2 items and word-initial in the M2–M3 items. These novel words were evaluated by 10 native Tibetan speakers (who did not participate in the subsequent experiment) to confirm that they had no actual meaning. They performed a binary real-word judgment, and there was 100% consensus that all items had no attested meaning. Subsequently, each novel word was embedded in six highly constraining sentences. Sentence constraint was established by expert judgment, consistent with conventions in related Chinese-language studies [73,74]. In highly constraining sentences, participants could infer the meaning of the novel word from the sentence context. This process generated eight sets of sentences, with each set containing six sentences. The morphemes, novel words, and sentences used in Experiment 2 were all different from those in Experiment 1. The procedure for creating our experimental materials was consistent with previous studies [73,74].

Apparatus.

Same as in Experiment 1.

Procedure.

Same as in Experiment 1.

Measures.

Same as in Experiment 1.