Glossa Psycholinguistics

2.1 Predictive L2 gender processing

Previous studies offer various definitions of prediction, which often differ from study to study. The present study adopts a restrictive view of prediction that has been commonly used by studies previously conducted on sentence processing (see Kim & Grüter, 2021; Kuperberg & Jaeger, 2016). In these studies, prediction is defined as a pre-activation of the upcoming nouns. This means that only effects observed before the onset of the critical words, in our case the target nouns, provide empirical evidence of predictive processing. Based on this assumption, previous eye tracking studies reported that language users can move their eyes to a particular object in the display even before the object is mentioned. Their relative proportion of looks to the objects is then interpreted as an index of how listeners predictively resolve reference, based on partial input as a sentence is unfolding (see Huettig et al., 2011). There is plenty of evidence that shows language users engaging in proactive, forward-looking processing during sentence comprehension (see Dahan et al., 2000; Foucart et al., 2014; Grüter et al., 2012; Hopp, 2016; Hopp & Lemmerth, 2018; Kaan et al., 2010; Lew-Williams & Fernald, 2010). As a result, prediction in language processing has been investigated in a wide range of linguistic domains: semantics (e.g., Casillas & Frank, 2013; Grüter et al., 2020; Kamide et al., 2003; Weber et al., 2016;), morpho-syntax (e.g., Hopp, 2016; Lemmerth & Hopp, 2019; Paorlieri et al., 2019) and discourse (e.g., Boudewyn et al., 2015; Kim & Grüter, 2021; Scheutz & Eberhard, 2004).

Previous studies show that native speakers of gendered languages use gender information on adjectives and varieties of determiners to anticipate upcoming nouns (e.g., Dussias et al., 2013; Halberstadt et al., 2018; Hopp, 2013, 2016; Hopp & Lemmerth, 2018; Paorlieri et al., 2010). It has been argued that the capacity to process gender predictively starts early in life for native speakers of gendered languages. Some previous studies also argued that simultaneous bilinguals are effective predictors, like monolingual speakers (Foucart et al., 2014; Grüter et al., 2012; Hopp, 2016; Kaan et al., 2010; Lew-Williams & Fernald, 2010) implying that early exposure to the second language scaffolds the mastery of predictive L2 gender processing. However, there have been inconsistent reports when it comes to the predictive behavior of late L2 learners. The earlier assumption was that late L2 learners have a Reduced Ability to Generate Expectation – RAGE (see Grüter & Rohde, 2013; Hawkins & Casillas, 2008). Predictive processing was conceived of as a capacity that is conditioned by the instantiation of the native language. According to this hypothesis, even advanced L2 learners may not be able to predict since they do not reach native-like automatization that is compulsory to predict. This argument has recently been challenged as the increasing number of studies reported predictive processing by both native and late L2 learners (see Bañón & Martin, 2021; Dussias et al., 2013; Fourcart et al., 2014; Hopp, 2013; Kaan, 2014; Kim & Grüter, 2021). These later studies contended that the same factors that influence L1 predictive processing also affect L2 processing. In other words, the competence to use prediction during L2 gender processing is constrained by the same individual-level factors (Bordag & Pechmann, 2007; Kim & Grüter, 2021; Schlenter, 2023). Therefore, L1-L2 difference in predictive behavior is attributable to modulating factors such as cross-linguistic influence, L2 fluency, memory load, and cue availability and reliability.

The discords associated with L2 predictive gender processing can be subsumed into two accounts of gender processing: processing-based and representational. The processing-based account associates the L2 predictive gender processing difficulty to the variability in the L2 lexical access and to various individual and language-level factors. It holds the view that L2 grammars do not differ from native grammars in terms of representation and that the variability in processing L2 morphosyntax predictively follows from difficulties in mapping the target morphophonological forms to syntactic features in situations of real-time processing pressure (see Prévost & White, 2000). Related to this, the Weaker Link Hypothesis (see Gollan et al., 2008) assumes that individual difference in gender assignment obscures the possibility of using gender for prediction since non-target lexical gender leads to a prediction error, implying that lexical gender and grammatical aspects of gender have a causal relationship. Similarly, the Non-selective Lexical Access Hypothesis (e.g., Salamoura & Williams, 2007) assumes that gender representations in the bilingual mental lexicon are interrelated or accessed non-selectively. Hence, L2 learners whose L1 marks grammatical gender typically show more target-like gender agreement processing than L2 learners whose L1s lack gender. According to this hypothesis, L2 learners whose L1s realize gender agreement may not be invariably native-like in processing because target predictive gender processing can be limited to certain gender markings. Similarly, the Missing Surface Inflection Hypothesis (Prévost & White, 2000) predicts that L2 inflectional variability stems from failures in retrieving inflectional forms under real-time processing pressure. L2 learners usually rely on default forms of gender because they fall short of mapping the appropriate target lexical form into the given grammatical context. The Shallow Structure Hypothesis (Clashen & Felser, 2006), Lexical Bottleneck Hypothesis (Hopp, 2018) and Lexical Gender Leaning Hypothesis (Grüter et al., 2012) make similar predictions. Put together, these models show that a strong lexical gender representation is a prerequisite for successful predictive processing of L2 gender agreement.

The representational accounts (Hawkins, 2009; Howard, 2011; Jakubowicz & Roulet, 2004) assume that there is an inherent difference between L1 and L2 learners that emerges from input type and learning strategies. For instance, during L1 acquisition, nouns are acquired in continuous speech, together with other constituents such as determiners and adjectives. However, during L2 learning, nouns are taught separately in a written form, in a way that disrupts the smooth association between nouns and other elements. Therefore, according to this account, L2 learners whose L1s do not encode grammatical gender do not project syntactic gender features for agreement and hence need to rely on non-grammatical cues for computing gender concord in the L2. For L2 learners, the predictive effects of gender marking could be limited to some memorized exceptions of listed noun-gender associations, because feature-based syntactic agreement between determiners and nouns is not available. This account also appeals to age-related grammatical impairments arguing that L2 grammars do not encode abstract syntactic gender features necessary for computing agreement relations unless these features are instantiated in the L1 (see Hawkins, 2009; Hawkins & Casillas, 2008).

2.2 Cross-linguistic influences and other constraints

Predictive L2 gender processing can be influenced by various factors. For example, it was argued that the availability of a higher number of gender cues in the input aids the use of gender for prediction (Fowler & Jackson, 2017; Garrido-Pozú, 2022). It was also indicated that training of gender assignment boosts the use of prediction for L2 gender processing (see Hopp, 2016). Mode of input presentation was also stated as one of the contributing factors. For instance, in the study conducted by Hopp & Lemmerth (2018) on cross-linguistic lexical and syntactic influence in grammatical gender, the elicited production task showed that all bilingual children assign target gender to nouns, irrespective of whether the nouns belong to the same or different gender class in German and Russian. In the visual world paradigm, however, only simultaneous bilinguals could predict. Simultaneous bilingual children showed predictive gender processing, irrespective of the gender (in)congruency between L1 and L2. In contrast, the successive bilingual children showed predictive gender processing only for lexically congruent nouns. Based on these findings, the study argued that the asynchronous acquisition of L2 in successive bilinguals implicates that L2 gender is first accessed through the L1 lexicon.

Moreover, studies reported that syntactic gender similarity between the L1 and L2 enhances predictive L2 gender processing (see Bordag & Pechmann, 2007; Bosch & Foppolo, 2022; Dussias et al., 2013; Morales et al., 2016). Conversely, a mismatch in gender agreement between the L1 and L2 poses difficulty for processing L2 gender agreement presumably since predictive gender processing requires easy and rapid access to grammatical gender knowledge (Dussias et al., 2013; Hopp, 2013; Lew-Williams & Fernald, 2007, 2010). For example, in successive experiments, Foucart and Frenck-Mestre (2011) investigated the effect of similarity between the first and the second language on grammatical gender processing in L2, focusing on German advanced learners of French and French native controls. The study explored gender agreement violations between determiner and noun, postposed adjective and noun, and preposed adjective and noun. The ERP results showed a similar P600 effect for native and non-native speakers for agreement violations when agreement rules were similar in the L1 and L2 whereas no effect was found for L2 learners when agreement rules varied across languages. The study concluded that syntactic processing in the L2 is affected by the similarity of syntactic rules in the L1 and L2. Recently, Bosch and Foppolo (2022) also investigated linguistic prediction based on grammatical gender in Italian-German bilingual children, using a visual world eye tracking paradigm. Children listened to sentences while looking at objects that either matched or mismatched in grammatical gender, and that varied with respect to cross-linguistic gender congruency. The study reported rapid predictive processing, as the children anticipated nouns based on the grammatical gender of the determiners. Furthermore, the children exhibited a gender congruency effect, i.e., they showed negative cross-linguistic influence when the grammatical gender of the two languages did not overlap, leading to delayed anticipation.

There is also evidence that, like syntactic overlap, lexical gender similarity between L1 and L2 plays a crucial role in L2 predictive gender processing. For example, Paolieri et al. (2020) investigated whether processing a word in one language is affected by the gender of its translation equivalent in another language. A group of Catalan–Spanish bilinguals performed a translation-recognition task while event-related potentials (ERPs) were recorded. The ERP data showed a reduced N400 for the congruent condition, implying facilitation between the lexical gender of the first and second languages. Moreover, Morales et al. (2016) investigated gender-congruency effects using the visual world paradigm. The study reported a reduced eye gaze latency in the congruent condition. Paolieri et al. (2019) also investigated the role of lexical gender in processing gender information. In the study, Russian–Spanish bilinguals showed gender congruency effects when they translate concrete nouns in isolation or in noun-phrases. Several other studies also reported lexical gender congruency effects (see Bobb et al., 2015; Dussias et al., 2013; Hopp, 2016; Hopp & Lemmerth, 2018).

The ability to predict during L2 gender processing can also be constrained by L2 fluency (see Blumenfeld et al., 2016; Hopp, 2016; Hopp & Lemmerth, 2018). However, the relationship between L2 fluency and L2 predictive gender processing is the complex one. Several studies previously reported greater or native-like prediction competency of speakers with advanced L2 fluency (see Chambers & Cooke, 2009; Dussias et al., 2013; Hopp, 2013; Hopp & Lemmerth, 2018). Equally, a substantial number of studies failed to replicate the L2 fluency effect (see Dijkgraaf et al., 2017; Hopp, 2015; Kim & Grüter, 2021; Mitsugi, 2020). Some other studies indicated that advanced L2 learners are different from both native and intermediate L2 learners in predictive use of gender cues (Dijkgraaf et al., 2019; Kaan et al., 2016). These show that the relationship between L2 fluency and L2 predictive gender processing is an area that awaits further investigation. The problem lies partly in measuring the fluency per se. Previous studies that examined the link between L2 fluency and L2 predictive gender processing used fluency tests that varied from perception test, receptive vocabulary to more standardized tests (see Segalowitz, 2016 and Tavakoli, 2016 regarding the difficulty in measuring L2 fluency).

Gender assignment errors is another factor that increases difficulties in computing gender agreement. If the target gender assignment has not been established, it usually leads to prediction errors. In turn, this affects the reliability of grammatical gender cues and hinders the use of prediction during L2 gender processing. Related to this, Hopp (2012) explored lexical and syntactic aspects of gender processing in L2 production and comprehension in advanced English learners of German and German native speakers through a picture naming task and a visual world paradigm. Results showed a strong correlation between variability in gender assignment during production and variability in using gender as a predictive cue in comprehension. Only L2 learners that had target-like performance in gender assignment in the production task were able to use gender as a predictive cue in comprehension. This finding shows that the variability in using gender predictively is more likely related to lower levels of activation and access to gender nodes rather than deficits of mental representation in the L2 grammar. Predictive L2 gender processing may also be influenced by non-linguistic factors such as motivation (Luna & Peracchio, 2002) and memory load (Gordon et al., 2001; Ito et al., 2018). To the best of our knowledge, not many studies have investigated the roles these factors play in L2 predictive gender processing.

2.3 Gender in Oromo and Amharic

Gender is often defined as classes of nouns that are reflected on other associated words (Hockett, 1958). The link between nouns and the associated words is expressed via gender agreement. Languages differ in the way they mark gender (see Corbett, 1991). Different dialects can also have different gender systems as is the case in Oromo. The Oromo gender system varies across dialects (see Clamons, 1992, 1993; Feleke & Lohndal, 2023; Feleke, 2024a). The Eastern dialect has a phonology-based gender assignment pattern, i.e., nouns that end in consonants and in the low central vowel /a/¹ are masculine whereas nouns that have other vowel /e, i, o, u/ endings are feminine (cf: (1a, c, e) and (1g)). In the dialect, the vowel endings serve as declension class markers (see Feleke, 2023; Feleke & Lohndal, 2023; Feleke, 2024a). Conversely, the Western dialect does not have grammatical gender (see (1b), (1d), (1f) & (1h)) since the feminine gender has been entirely neutralized (see Clamons, 1992, 1993; Feleke & Lohndal, 2023; Feleke, 2024a). Therefore, in the Eastern dialect, nouns agree in gender with elements such as adjectives, verbs, and determiners. In the Western dialect, however, this agreement does not exist due to the neutralized feminine gender.

Amharic has a default masculine gender (see Kramer, 2015). Most Amharic nouns are masculine in gender, but Amharic has maintained the masculine-feminine gender distinction. Hence, the Eastern Oromo dialect patterns with Amharic in terms of gender agreement. For example, in both the Eastern dialect and Amharic, nouns agree in gender with verbs ((2c), (2d), (2g) & (2h)), adjectives ((2c) & (2d)) and determiners ((2a), (2b), (2e) & (2f)).

In Amharic, only definite nouns agree in gender with attributive adjectives, i.e., there is no gender agreement between indefinite Amharic nouns and attributive adjectives. Predicative adjectives never agree in gender with nouns in Amharic (see (2b)). Moreover, when there is a sequence of adjectives, only the first adjective necessarily agrees in gender with noun in Amharic; the agreement between the noun and the rest of the adjectives is optional (see Kramer, 2015). In the Eastern Oromo dialect, adjectives always agree in gender with the noun³. In the dialect, only the proximal demonstrative agrees in gender with the noun, and the distal demonstrative never agrees in gender with the noun. In Amharic, however, both distal and proximal demonstratives agree in gender with the noun. Besides, Amharic adjectives are strictly prenominal tɨllɨk bet ‘a big house’. Interrogative pronouns have a flexible position yətu bet new? or bet yətu new? ‘Which is a house?’. Demonstrative pronouns also have a flexible position in Amharic: yɨh bet new or bet yɨh new ‘This is a house’, but the postnominal position is used in very rare contexts, for example when a special emphasis is placed on the noun.

Table 1 presents examples of the three gender agreement domains that we investigate: noun-adjective, noun-interrogative pronoun, and noun-demonstrative pronoun gender agreement. Examples (3a) and (3b) show that, in the Eastern Oromo dialect, the feminine noun abaaboo and the masculine noun mana agree in gender with the adjective – guddoo and guddaa respectively. In the Eastern dialect, there are also distinct feminine tamii and masculine kamii interrogative pronouns. Likewise, there are distinct masculine and feminine proximal demonstrative pronouns – tun and kun respectively. However, in the Western Oromo dialect ((3c) & (3d)), only the masculine adjective guddaa is acceptable, and the feminine counterpart guddoo does not exist. Moreover, in the Western Oromo dialect, there is only one interrogative pronoun, kami, and one proximal demonstrative pronoun kun since the feminine counterparts have been neutralized. Amharic gender agreement (see (3e) & (3f)) patterns with the gender agreement of the Eastern Oromo dialect. Hence, in Amharic, the feminine noun s’ɨggerəda and masculine noun bet agree in gender with the adjective – tɨllɨk’wa, and tɨllɨk’u respectively. Moreover, Amharic has distinct feminine yətwa and masculine yətu interrogative pronouns. Likewise, it has distinct masculine yɨh and feminine yɨčči proximal demonstrative pronouns.

From a sociolinguistic standpoint, Oromo is spoken in the Oromia Regional State, and it serves as the working language of the region. Furthermore, Oromo serves as a medium of instruction in the schools across the region. The Western dialect is the dominant school lect, but there is a high degree of flexibility regarding dialect use in the schools; for example, the schoolteachers sometimes mix their local dialect during classroom instruction (see Feleke, 2024a). In general terms, most Oromo communities except the speakers of the Western dialect, are bidialectal speakers since they use a different dialect in the school and other contexts. As a native lect, the Eastern dialect is spoken in the eastern part of the Oromia Regional State whereas the Western dialect is spoken in the western part of the Region. Native speakers of the Eastern dialect are exposed to the Western dialect via schooling (see Feleke & Lohndal, 2023; Feleke, 2024a). However, native speakers of the Western dialect do not have exposure to the Eastern dialect. Amharic is the working language of the Federal Government of Ethiopia. It is spoken as a first language in the Amhara Regional State and in most cities and towns across the country (Meyer, 2006). Amharic is also spoken as a second language almost all over the country, and it is taught as a second language in the schools in Ethiopia. The grade level in which Amharic is introduced varies from state to state. In the Oromia region, which is the target of the present study, Amharic is introduced as a subject at grade five⁵. This means that L1 Western L2 Amharic speakers must acquire gender marking in schools after the exposure to Amharic since their native Western dialect does not encode grammatical gender. This is not the case for the speakers of the Eastern Oromo dialect because they have prior exposure to a gender system that is compatible with the gender system of Amharic.

4.1 The Visual World Paradigm (VWP)

We used the webcam-based visual world paradigm as an alternative to infrared eye tracking (see Bott et al., 2017; Prystauka et al., 2023; Vos et al., 2022; Yang & Krajbich, 2021). Using Gorilla Experiment Builder, the participants were presented with a visual display showing two pictures, either in the same or in the different gender condition. In the same gender condition, the target and competitor pictures have the same gender and in the different gender condition, the target and the competitor pictures have different gender. Each target-competitor pair was presented along with an audio instruction. The participants selected the target picture based on the instruction. Meanwhile, their eye movements were recorded. The task was administered in a temporary laboratory that we established at Haramaya University.

4.1.2 Stimuli and procedures

4.1.2.1 The stimuli

Sixteen (16) pictures of inanimate objects were selected from electronic and print sources. Four of the nouns designated by the pictures have masculine gender both in the Eastern Oromo dialect and in Amharic (M-M), four nouns have feminine gender both in the Eastern dialect and in Amharic (F-F), four nouns have masculine gender in the Eastern dialect but feminine gender in Amharic (M-F), and the remaining four have feminine gender in the Eastern dialect, but masculine gender in Amharic (F-M). Hence, half of the pictures designate gender congruent nouns and the remaining half gender incongruent nouns. The frequency of the nouns designated by the pictures were matched across the congruent and incongruent conditions, using SKELL/SKETCH ENGINE⁶ (Kilgarriff et al., 2015). There was no significant difference between the frequency of the congruent nouns and the incongruent nouns, independent sample t-test, t = .743, p = .463. During the test design, the gender congruent pictures were paired with each other, and the gender incongruent pictures were always paired with the gender incongruent pictures. Therefore, in the gender congruent condition, both the target and the competitor pictures have the same gender in the Eastern dialect and in Amharic. Conversely, in the gender incongruent condition, both the target and the competitor pictures have different gender in the Eastern dialect and in Amharic. The 16 pictures were presented in a 2*2*3 design: two gender conditions (same, different), two stimuli conditions (target, competitor), and three gender agreement conditions (noun-interrogative, noun-adjective, noun-demonstrative). These resulted in 192 possible target-competitor combinations, and total of 96 test items ((16*2*2*3)/2). In each item, the target picture appeared either to the right or to the left of the computer screen. Each target-competitor picture pair was presented in the same color. There were additional 14 distractors. In the distractor condition, pairs of pictures of numerals were displayed on the computer screen with audio instructions such as ‘Among the following, which is number one?’. In sum, there were 110 test items, 96 experimental and 14 distractor items. The items were randomized and equally divided into two in such a way that the gender agreement domains and distractor items are equally distributed between the categories. Half of the test items (55) was administered to half of the participants in each study group, and the remaining half was administered to the rest of the participants. The pictures were equally sized into 400 by 400 pixels.

Table 4 presents samples of the Amharic audio instructions. The audio instructions were recorded by the adult female native speaker of the Standard Amharic (age 28). We provided the adult speaker with written list of Amharic sentences and asked her to read the sentences with natural pace and accent. We segmented the recorded sentences using Audacity. We used silence to align the gap between the segments. The recorded instructions contain the three gender agreement conditions. In the noun-interrogative pronoun condition ((4a) & (4b)), the cues are gender suffixes attached to the interrogative pronoun. In the noun-adjective condition ((4c) & (4d)), the cues are gender suffixes on the first adjectives, and in the noun-demonstrative pronoun condition ((4e) & (4f)), the cues are gender suffixes attached to the demonstrative pronoun. In all conditions, we were interested in the time window between the onset of the words that carry the gender cues and the target nouns. In each condition, there was a gender-neutral adjective which served as a buffer between the gender cues and the target nouns. In Section 2.3, we showed that in Amharic, only definite nouns agree in gender with adjectives. Besides, when there is a sequence of adjectives in a sentence, only the first adjective obligatorily agrees in gender with noun.

Table 4: Sample Amharic audio instructions.

No.	Cond.	G	Carrier phrase	Cues	Buffer	Targets	Post-targets	Gloss (condensed)
(4a)	N-INT	f	kǝnnǝzzih wɨst’	yǝt-wa	k’əyy	s’ɨggerəda	ɨndǝhonǝčč asayy	…which one is a red rose?
(4b)	N-INT	m	kǝnnǝzzih wɨst	yǝt-u	k’əyy	c’amma	ɨndǝhonǝ asayy	…which one is a red shoe?
(4c)	N-Adj	f	kǝnnǝzzih wɨst’	k’əyy -wa	k’onğo	s’ɨggerəda	yǝt ɨndallǝčč asayy	Show where a nice red rose is
(4d)	N-Adj	m	kǝnnǝzzih wɨst’	k’əyy -u	k’onğo	c’amma	yǝt ɨndallǝ asayy	Show where a nice red shoe is
(4e)	N-DEM	f	kǝnnǝzzih wɨst’	yɨhččin	k’əyy	s’ɨggerəda	bǝtɨkɨkkɨl assay	Select this red rose
(4f)	N-DEM	m	kǝnnǝzzih wɨst’	yɨhɨn	k’əyy	c’amma	bǝtikɨkkɨl assay	Select this red shoe
			1700ms	1100ms	900ms
				Target window

The audio stimuli were sequenced as follows: the duration between the onset of the carrier phrases and the words that carry the gender cues was 1700ms. The duration from the onset of the carrier phrases to the onset of the buffers was 2800ms (1700+1100ms). The duration from the onset of the carrier phrases to the onset of the target nouns was 3700ms (2800ms +900ms). The critical time window for the anticipatory processing extends from the onset of the words that carries gender cues to the onset of the target nouns (1100 + 900 = 2000).

4.1.2.2 Procedures

The experiment was programmed in the Gorilla Experiment Builder platform (Anwyl-Irvine et al., 2020). We used the Gorilla platform which employs the recent version of WebGazer (available since November 11, 2021). The study started with written instructions explaining the purpose and the general procedure of the experiment. Then, the participants were directed to the eye-tracking task. The task consisted of 2000ms preview time followed by a fixation cross. The fixation cross was displayed for 500ms. Then, the eye tracking window appeared in which each pair of the target and competitor pictures was displayed along with the audio instruction. The pair of the pictures was displayed (see Figure 1) indefinitely until one of the pictures was selected by the participants, using the mouses-click. The experiment was conducted on 4 PCs with the monitor size of 1366 × 768, 1280 × 720, 1024 × 768, and 1536 × 864 pixels. Maximum of four participants were tested simultaneously, wearing headphones.

Figure 1: The sequence of stimuli displays in the visual world paradigm, s’ɨggerəda ‘rose’ in the left and c’amma ‘shoe’ in the right.

The participants were instructed to choose the pictures based on the audio instructions. They were also told that the task would last approximately 25 minutes. They were not instructed to sit at a particular distance from the screen, but before the experiment it was assured that the laboratory had sufficient light. The experiment began with a practice session of 10 items from different materials. Then, the participants were directed through a calibration phase of 5 points presented consecutively across the entire screen. They were instructed to look at and click on each point which was visible for 3 seconds. They were told to look at the points without moving their heads. Each participant had to pass the calibration, which we set to 60% (3/5) of the gaze prediction. Following the validation, the participants received visual feedback (plots) on their performance. If 60% of the threshold was reached for the validation plot, the participants could proceed to the experiment. If not, they were looped back to the start of the calibration phase. If the calibration was not successful within 5 attempts, the study was aborted. The calibration was repeated once half-way in the experiment, after 28 trials, to compensate for the small head movements. The total study duration of the experiment was around 25 minutes.

4.2 The picture description production task

After the eye-tracking task, participants performed a gender assignment task. The task was used to determine the participants’ knowledge of L2 lexical gender. The same 16 pictures from the eye-tracking experiment were used in the task. Each picture, for example rose, appeared at the center of the computer screen. Then, the participants were asked to describe orally the advantage of the object designated by picture, using a complete Amharic sentence. After seeing the picture of rose, for example, the participants produced sentences such as in (5).

The gender assigned to the target noun (rose) was determined based on the gender agreement that was established between the noun rose and the verb serve. Since Amharic allows gender agreement between nouns and verbs, we exploited this agreement relationship to probe into the participants’ knowledge of L2 gender assignment. The participants were given a score of 1 for the correct agreement and 0 otherwise. The L2 gender assignment score is the number of responses that contain the correct gender agreement between the noun and the verb.

4.3 Preprocessing steps

The Gorilla engine provides raw and normalized coordinates of the eye fixation, and we used the normalized coordinates in our analyses. Gorilla also provides face-conf, which is a data quality metrics. Face-conf indicates how strongly the image under the model resembles a face. Its values vary from 0 to 1, and values greater than 0.5 are indicative of a good model fit. We excluded values less than 0.6 on the face-conf metrics which together constituted 8 (4.4%) participants. The sampling rate of our original sample (N = 180) varied from 3.88Hz to 59.42 Hz (M = 36Hz). We excluded additional two participants whose sampling rate was less 5Hz. In total, 10 (5.5%) participants were excluded. The average sampling rate in the resulting groups (170 participants) was 37Hz. This means that, in average, 37 data points were sampled every 1000ms, and one data point was acquired every 27ms.

Our analyses were based on the proportion of eye fixations within the preselected time window, 2000ms from the onset of the word that carries the gender cues to the onset of the target nouns. While determining the cues onsets, we included 200ms extra time to compensate for the time taken to launch the eye movement. Then, we performed two-step statistical analyses – first on the time course of prediction, then on the factors that modulate predictive L2 processing. We inspected the time course of prediction using divergence point analysis (Stone et al., 2021) and cluster-based permutation test (Ito & Knoeferle, 2023; Maris & Oostenveld, 2007; Meyer et al., 2021). In the divergence point analysis, the language group was independent variable, and the dependent variable was the proportion of looks to the target and competitor pictures. We used non-parametric bootstrapping divergence point in which the existing dataset is resampled multiple times to generate “new” datasets, and a statistical test is applied after each resample. The resampling was stratified by subject, timepoint, and stimuli conditions (target and competitor). A new divergence point was estimated after each resampling. Then, a distribution of divergence points was generated whose mean was taken as the overall divergence point. We bootstrapped means and confidence intervals (CIs) to determine the significant point of divergence.

For further condition-specific fine-grained analysis, we performed cluster-based permutation test following the manual provided by Ito and Knoeferle (2023). In this case, the agreement conditions (noun-adjective, noun-demonstrative and noun-interrogative) and language groups were independent variables, and the proportion of looks to the target and competitor pictures was the dependent variable. Cluster-based permutation analysis (CPA) is also a non-parametric test, and it detects the difference between two conditions in a preselected time window – 2000ms in our case. A significant result indicates a robust effect of condition. We performed the cluster analysis based on LME, using the clusterperm.lmer function in the permutes R package. In both divergence and cluster-based analyses, we grouped the timepoints into 30ms bins because one data point was acquired approximately every 27ms which can be rounded to 30ms.

The analysis of the time course of prediction was followed by the analysis of the factors that modulate the predictive processing. For the latter analysis, we performed logistic mixed effects regression to detect the effect of the gender conditions (same and different), language groups (L1 Amharic, L1 Eastern and L1 Western), knowledge of lexical gender, and L2 fluency on the proportion of looks to the target pictures. We performed a similar but separate analysis for the congruency effect. The separate analysis was necessary since gender congruency could be established only between the Eastern Oromo dialect and Amharic. For the regression analysis, we grouped the timepoints into 50ms bins since we wanted to reduce the potential effect of autocorrelation by grouping the timepoints into larger bins. To reduce noise in the data, both time course and factor analyses were performed only on nouns whose lexical gender was properly assigned by the participants. We also excluded eye fixations outside the participants’ screen dimension. The analyses were conducted in R (version 4.2.3, 2023).

1.1 Predictive processing and the role of differential cross-linguistic influence

We have seen that the L1 Eastern dialect speakers have earlier onset of predictive looks than the L1 Western dialect sparkers, implying that cross-linguistic influence is a crucial factor in L2 predictive gender processing. We can take this as evidence of facilitation from the Eastern Oromo dialect. Several previous studies recognized the pivotal role that cross-linguistic influence from the first language plays in predictive L2 gender processing (e.g., Bosch & Foppolo, 2022; Dussias et al., 2013; Garrido-Pozú, 2022; Hopp, 2016; Hopp & Lemmerth, 2018; Morales et al., 2016). Data from the Oromo-Amharic bilingual speakers further elucidate that morphosyntactic similarity between the Eastern Oromo dialect and Amharic not just enhances the predictive processing of Amharic gender agreement, but also enforces predictive processing. Quantitatively, the L1 Eastern dialect speakers predicted faster than the L1 Amharic speakers. Such a strong enforcement of grammars of the first language was previously reported by Meir et al. (2020) in the study they conducted on the processing of case marking by Russian-Hebrew bilingual speakers. In their study, Russian-Hebrew bilingual speakers exploited Hebrew case markers more effectively than the monolingual Hebrew speakers. The strong enforcement during the L2 case processing was associated with the robust case marking in Russian as opposed to the sporadic case marking in Hebrew. When it comes to the Oromo-Amharic bilingual speakers, we have seen that in Amharic, gender is assigned based on semantic parameters. In contrast, the Eastern Oromo dialect has robust phonological endings or declension classes that make the gender of the nouns unambiguously identifiable. The Eastern Oromo dialect also has a richer gender morphology than Amharic. For example, only attributive adjectives agree with nouns in gender in Amharic. However, in Oromo, every adjective agrees with noun in gender.

We have seen that lexical gender congruency between Amharic and the Eastern Oromo dialect does not significantly affect predictive L2 gender processing. This means that lexical gender congruency does not conspire with the syntactic gender congruency in the process of facilitating predictive L2 gender processing. Many previous studies argue that lexical gender congruency plays a crucial role in predictive L2 gender processing (e.g., Hopp, 2016; Lago et al., 2021; Morales et al., 2016; Paolieri et al., 2020). The Oromo-Amharic bilingual data do not replicate this claim. As we indicated in the introduction and elsewhere, speakers of the Eastern Oromo dialect are exposed to the neutralized Western Oromo dialect. We suspect that the facilitative role of the lexical gender congruency between Amharic and the Eastern Oromo dialect is probably suppressed by the speakers’ exposure to the neutralized Western Oromo dialect.

1.2 The role of L2 fluency and knowledge of L2 lexical gender

In the above discussions, we argued that cross-linguistic influence is the most important factor that mediates L2 predictive gender processing. However, cross-linguistic influence is not the only important factor. Knowledge of L2 lexical gender influences L2 predictive gender processing too. We have seen that knowledge of lexical gender interacts with the gender condition. In other words, speakers that have better knowledge of lexical gender frequently gaze at the target pictures in the predictable different gender condition than speakers that have little knowledge of lexical gender. This clearly shows that there is a link between the mastery of L2 lexical gender and L2 predictive gender processing. This is not the first time this has been found. Several previous studies recognized the importance of the knowledge of L2 lexical gender in predictive L2 gender processing (see Clashen & Felser, 2006; Grüter et al., 2012; Hopp, 2013, 2016; Lemmerth & Hopp, 2018, 2019; Prévost & White, 2000). Indeed, most lexicalist models of L2 gender processing such as the Shallow Structure Hypothesis (Clashen & Felser, 2006), the Lexical Bottleneck Hypothesis (Hopp, 2018), the Missing Surface Inflection Hypothesis (Prévost & White, 2000), and the Lexical Gender Learning Hypothesis (Grüter et al., 2012) heavily rely on the integration between lexical and syntactic representations. As it stands, it seems that the knowledge of L2 lexical gender contributes to L2 predictive gender processing. In our case, it is a secondary factor of importance after differential cross-linguistic influence.

We have also seen that there is a strong association between L2 fluency and predictive L2 gender processing. Speakers that have advanced Amharic fluency frequently gaze at the target pictures in the predictable different gender condition than speakers that have lower Amharic fluency. The evidence suggests that a native-like fluency in L2 leads to an automated processing of L2 gender agreement. The result strengthens previous studies that reported a positive association between L2 fluency and L2 predictive gender processing (e.g., Blumenfeld et al., 2016; Chambers & Cooke, 2009; Dussias et al., 2013; Hopp, 2013; Hopp, 2016; Hopp & Lemmerth, 2018).

More importantly, from the results of the cluster-based permutation analysis we have learned that L1 Eastern dialect speakers effectively employ the gender cues to anticipate the upcoming nouns in all agreement conditions. The inconsistency in the noun-pronominal order does not affect the predictive processing of the L1 Eastern dialect speakers. In Amharic, interrogative pronouns and demonstrative pronouns have flexible word order in a sentence. This flexibility in the position of the gender cues only influences the predictive processing of the L1Western dialect speakers; the L1 Western dialect speakers do not predict in the most flexible noun-interrogative pronoun gender agreement condition. The result partly coincides with previous studies that stress the importance of cue availability or reliability (Falk & Bardel, 2011; Garrido-Pozu, 2022; Grüter & Rohde, 2013; Hawkins, 2009; Jakubowicz & Roulet, 2004; Kaan, 2014; Kaan & Grüter, 2021). Meanwhile, it also indicates that L2 learners can withstand input variability in the L2 if they are native speakers of languages or dialects that have rich gender morphology. It could be that, once the gender system of the first language is instantiated in the early childhood, cue availability or reliability ceases to become an essential factor in predictive L2 processing.

In general, data obtained from the Oromo-Amharic bilingual speakers imply that bilingual speakers predict only in a conducive condition, i.e., in the condition that is attenuated by facilitation from the first language, knowledge of L2 lexical gender, L2 fluency, and the availability of the gender cues. Some previous studies also argue that predictive processing is deployed only in a conducive condition (see Schlenter, 2023). By implication, prediction during L2 gender processing is optional; it aids the processing automatization only if the condition is favorable. In the unfavorable condition, learners may rely on other cues such as semantic relationship (see Prévost & White, 2000). Apparently, the joint contribution of several factors creates a suitable condition that fosters the deployment of prediction during L2 gender processing. The optionality of prediction during L2 gender processing stands against previous representation-based models of L2 predictive processing such as the Reduced Ability to Generate Expectation (RAGE) (Grüter & Rohde, 2013) and the Representational Deficit Hypothesis (Hawkins, 2009). These models assume that L2 learners whose first languages do not encode grammatical gender may fail to acquire the capacity to predict. Data from the Oromo-Amharic speakers confirm that the lack of prediction is not an inherent trait of L2 learners. L1 Western L2 Amharic speakers whose first language or dialect does not encode grammatical gender could employ gender cues to predict, given that the condition is conducive. If there exists an inherent deficit associated with L2 prediction, this would not have been possible. Rather, the results indicate that whether to predict or not during L2 gender processing is largely determined by language-specific and individual level factors. In this regard, we illustrated that L1 Western dialect speakers may or may not be able to predict depending on their level of L2 fluency, knowledge of L2 lexical gender, and the availability of the gender cues. Many studies have previously arrived at the same conclusion but based on well-studied languages (see Kaan, 2014; Schlenter, 2023).

A.1. Effects of language group, knowledge of lexical gender and L2 fluency

A.2. The Effect of gender congruency