AWARENESS AND LEARNING FORM-MEANING CONNECTIONS

If it is the case that learning depends on awareness at the level of noticing the relevant stimuli, then in the case of learning form-meaning connections, the implication is that it is necessary to attend to both form and meaning. The first time a word is encountered, the learner might only be able to guess at its meaning or encode a rough approximation with the help of a teacher or dictionary. This would constitute a working hypothesis as to the meaning of the word. What is critical here is that not only are form and meaning attended to, but that they are also integrated by a form-meaning connection. In all but the simplest cases, the true meaning of the word would need to be derived by abstracting over a number of form-meaning pairings. This could be achieved by either implicit or explicit means. In the explicit case, when the word is encountered again, the learner might explicitly recall the previously hypothesized meaning. If this does not seem to fit the current context, the hypothesis could be modified. Through a cyclic process of explicit hypothesis formation and testing, the learner might come to acquire the meaning of the word. However, it is also possible that the true meaning of a word is abstracted through implicit learning mechanisms. This process might be no more than a consequence of the way in which instances of form-meaning connections are stored in memory, as proposed by connectionist memory models (McClelland & Rumelhart, 1985).

Many researchers of both first (L1) and second (L2) language acquisition stress the role of attention in learning form-meaning connections. In L1 acquisition, Merriman (1999) argued that the child must attend to a form in relation to a certain aspect of meaning. The task for the researcher is to determine the factors that influence the child's distribution of attention, such as the shape bias (Smith, 1999) or the search for distinctive features (Merriman). For example, when faced by a second label for an object, the principle of mutual exclusivity is violated and the child is assumed to “seek out, discover, and dwell on features that are uniquely associated with the second label” (Merriman, p. 350). This sounds very much like an explicit process of hypothesis formation and testing.

With regard to L2 acquisition, Doughty and Williams (1998) stressed not only attention to form and meaning but also their integration within the learning episode. It is not enough that the relevant form and meaning are attended to; the learner must actually hypothesize a connection between them. This is an elaboration of the concept of noticing within the context of learning form-meaning connections. N. Ellis (1994) went further and argued that learning form-meaning connections is dependent on declarative memory—the same memory systems that are responsible for the storage of personally experienced events, broadly referred to as episodic memory. Remembering that you think that the word ravenous means something like “very hungry” is therefore like remembering that you ate a particularly large breakfast last Sunday morning, and working out the precise meaning of ravenous involves explicit recall of contexts of use of the word.

N. Ellis' (1994) arguments are based on the fact that variables that are known to affect episodic memory also influence vocabulary learning (e.g., intelligence, depth of processing, and metacognitive and mediational strategies). Gupta and Dell (1999) made a similar argument on purely computational grounds, pointing out that the arbitrariness of form-meaning mappings makes them very difficult to assimilate using the kinds of implicit (in their view, connectionist) learning mechanism that underlie skill learning. Important evidence is also provided by research on amnesia. For example, amnesics, such as H. M. (Gabrieli, Cohen, & Corkin, 1988) or SS (Verfaellie, Koseff, & Alexander, 2000), can learn novel word forms, as assessed by tests of implicit memory such as fragment completion and priming. However, they seem to have great difficulty learning new vocabulary; for example, they are unable to recognize or use words that they have been exposed to over long periods after the onset of amnesia. Amnesics are unable to recall specific experiences; they have impaired episodic memory. Their inability to acquire vocabulary thus implies that episodic memory is involved in learning form-meaning mappings.

However, more recent work suggests that amnesics can learn vocabulary after all. Vargha-Khadem et al. (1997) examined three young adults who had suffered brain damage in infancy and become diagnosed as amnesics during their school years. Amazingly, their academic progress is reported to be in the normal range. Despite having great difficulty in recalling a story after a short delay, recounting a day's events, or remembering when their lessons were, they showed normal vocabulary and reading development and general knowledge. Other studies have also provided evidence that even severe adult amnesics can acquire vocabulary and factual information. Kitchener, Hodges, and McCarthy (1998) reported on a patient who could not remember highly salient personal events (e.g., that his daughter had left home or that his mother had died), and yet he was able to distinguish real and fictitious public events and to choose the correct definitions of words that came into the language only after the onset of his amnesia. Verfaellie et al. (2000) described a severely amnesic patient, PS, who showed preserved semantic learning, including learning of novel vocabulary. These studies count against the claim that vocabulary learning is necessarily dependent on memory for specific episodes of word use.¹

The consensus, therefore, appears to be that learning form-meaning mappings requires attention to form and meaning and probably their integration at encoding; that is, there needs to be awareness at the level of noticing of a form-meaning connection. However, there is less certainty over whether explicit, episodic memory is necessarily involved in working out word meanings from instances of use or in demonstrating knowledge of word meanings in comprehension and production (i.e., whether there needs to be awareness at the level of understanding).

PREVIOUS EXPERIMENTAL RESEARCH

Very little implicit learning research has specifically examined form-meaning connections. A notable exception is a study by DeKeyser (1995) that employed a miniature artificial language with rich inflectional morphology for marking biological gender, number, and case. For example, if the sentence subject was plural, the verb carried the suffix -it, and if the subject was singular, the verb carried the suffix -at. The training task was sentence-picture matching; participants had to indicate whether a given sentence correctly described a picture. Crucially, mismatching pairs always involved vocabulary differences and never purely grammatical errors, so the training task did not encourage participants to attend to the critical form-meaning relationships. Learning was measured by a production task in which participants were presented with a picture and had to type the correct sentence in the artificial language. Performance was at chance in the critical test of generalization to new sentences, despite the fact that the participants had endured 20 learning sessions of 25 minutes each. Thus, there was no evidence of implicit learning even of a simple subject-verb number agreement rule, even though that rule was very similar to one in the participants' L1, English. Note that participants showed good test performance for forms that had occurred during training, so they had obviously been paying sufficient attention to encode the relevant forms in memory during the training task. However, the training task did not require them to attend to the relevant aspect of meaning nor its potential relationship to forms. The results therefore support the idea that attention to form and meaning is necessary for learning.

One can imagine a situation in which both the relevant form and the relevant meaning are attended to, but the participant does not integrate them at encoding. In this case, the form and meaning can be said to be noticed, but the connection between them is not. Such a situation is clearly difficult to engineer, but in a previous work (Williams, 2004, Experiment 1), I reported a study that attempted to do just this. The system comprised eight novel nouns learned as translation equivalents of common English words (e.g., johombe/i “monkey/s”, nawase/i “vase/s”) and eight articles that were learned as translation equivalents of English “the” (ig, i, ga, ge), “a” (ul, ula), and “some” (tei, tegge). The participants were not informed why there were four words for “the” and two each for “a” and “some.” During training, the participants were presented with noun phrases formed by combining nouns and articles (e.g., i johombi, ga nawase, ul johombe, ge nawasi). For each phrase, they had to (a) repeat it aloud (to facilitate memory encoding), (b) indicate whether the noun referred to a living or nonliving thing, and (c) translate the phrase into English. The distribution of the articles was in fact governed by noun definiteness, number, and animacy (animates took ig, i, ul, tei; inanimates took ga, ge, ula, tegge). The training task explicitly drew the participants' attention to the animacy of the noun without hinting that this was relevant to the choice of article. During training, each noun appeared with only three of the possible four determiners licensed by the grammar, the fourth being withheld for testing generalization ability. The training period consisted of 15 repetitions of each noun phrase, each repetition occurring in a separate block of trials. This was followed by a test phase in which the participants had to translate English phrases (e.g., “the monkey”). They were presented with a choice between two possible noun phrases (e.g., ig johombe / ga johombe) where the articles were of the correct number and definiteness, but one violated the animacy rule. The critical test items concerned the noun phrases that had been withheld from the training phase, referred to as generalization test items. There were also items that had occurred during training, referred to as trained items. After performing this test, the participants were interviewed to ascertain their level of awareness of the system and how they had approached the task.

Somewhat surprisingly, even after the test phase, 30 out of the 37 participants were unaware that animacy was relevant to determiner usage—that is, they had no awareness at the level of understanding. However, their responses to the generalization test items were 61% correct, which was significantly above the chance level of 50%. It was also found that performance was correlated with the participants' language background. For the participants who spoke an L1 that encodes grammatical gender (referred to here as a gender language), generalization test performance was significantly better than for the participants whose L1 did not encode gender. Because there was also a correlation with overall number of gender languages known, r = 0.572, p < .001, it is not clear whether it was knowing a gender language as L1 or knowing many gender languages in general that was important. There was also a correlation between generalization test performance and phonological short-term memory (PSTM), r = .50, p < .01. However, given that PSTM was also correlated with number of gender languages, r = .519, p < .01, it is not possible to say whether the relevant underlying factor was memory ability or linguistic knowledge.

The previous experiment suggested that it is possible, in principle, to implicitly learn form-meaning connections when they are attended—but not integrated—at encoding. This is consistent with work in experimental psychology that has looked at learning associations under conditions where the stimuli are attended to but the participants are not aware of the relationship between them (Jiménez & Méndez, 1999; Logan & Etherton, 1994). The results count against the suggestion that learning form-meaning connections necessarily involves intentionally integrating form and meaning at encoding (Doughty & Williams, 1998). However, the correlation with knowledge of gender languages suggests that the situation might be more complex than this and that interactions with prior linguistic knowledge might also be important.

In Experiment 2 of Williams (2004), I attempted to reduce the potential impact of prior knowledge of gender languages by employing a system that was superficially less similar to article systems in languages that the participants were likely to know. The participants first learned that the novel words gi and ro meant “near” and the words ul and ne meant “far” (colors were used to distinguish the alternative forms for each meaning). It was then explained that these words functioned like the English definite article but also expressed how far away the object was. The participants were then provided with phrases such as ul dog and were told to construct sentence contexts that highlighted the distance relationship implied by the novel word (e.g., I threw a ball to ul dog that was at the bottom of the garden). As in the first experiment, there was in fact a correlation between determiners and noun animacy. Animate objects took gi and ul and inanimate objects took ro and ne. To force attention to the animacy distinction, the participants were instructed to construct different kinds of context for living and nonliving things. Living things were to be described in outdoor contexts, and nonliving things were to be described in indoor contexts. Thus, for ro telephone, an appropriate sentence would be “I was woken by ro telephone by my bed.” To help the participants focus on the relevant constraints during sentence production, they used icons on the computer screen to represent the noun (a picture of the object), near or far (using geometric figures), and indoor versus outdoor (a picture of a house or a tree). There were six nouns per class. Two in each class occurred with both possible determiners during training, the remaining four only occurred with one determiner, and the withheld items were reserved for testing generalization ability.

Participants performed this task for between 128 and 160 trials, depending on time constraints. There was then a test phase in which figurative representations of training and generalization items were presented (e.g., a picture of a lion, the far symbol, and a tree) along with a choice between two noun phrases (e.g., ul lion vs. ne lion). The participants were instructed to choose the phrase that seemed more familiar or “better.” After completing this task, only 3 of 17 participants tested expressed any awareness of the relevance of animacy (or the inside/outside distinction). Performance on generalization test items for the remaining 14 was not significantly different from chance (51.5%), whereas performance on test items that corresponded to trained items was highly accurate (81%). There were no correlations with language background factors. Thus, in contrast to Experiment 1 of Williams (2004), there was no evidence of implicit learning of form-meaning connections, even though the relevant forms and meanings were attended during the training task.

There are two potential problems with this experiment, however. First, the training task was rather arduous for the participants. It is known that—at least in some cases—implicit learning is reduced under dual-task conditions (for discussion in an applied linguistic context, see Carr & Curran, 1994). Therefore, overall task demands might have been an important factor. Second, the target system was rather small. There were only six nouns per category, and only four of these occurred with both possible determiners during training. It has been argued that a critical mass of exemplars might need to be encoded before generalizations can emerge (Marchman & Bates, 1994; Robinson & Mervis, 1998). The first experiment reported in the following section was therefore designed to address these problems.

A second problem is that the training tasks were highly artificial because of the requirement to explicitly draw the participants' attention to the relevant meaning feature, but not its relationship to the determiners. In the present experiments, a more natural situation was constructed in which participants merely listened to sentences containing the determiner-noun combinations (rather than generating them), indicated whether the novel word meant near or far, and formed an image of the general situation described by the sentence in anticipation of a test of memory for overall understanding of the sentence (not exact wording). Their attention was not explicitly drawn to the animacy feature at all; rather, this was implicit in the representation of the context. This task is less stressful and demanding than the sentence generation task. However, it also takes us one step further away from the noticing hypothesis. It seems unlikely that participants would spontaneously notice whether the relevant noun was animate or inanimate in this task. The task only requires them to notice the identity of the novel determiner. Any evidence of implicit learning of the form-meaning connection under conditions where only the form is attended would run counter to the prevailing emphasis on attention to form and meaning—and their integration—in discussions of learning form-meaning connections in the L1 and L2 learning literature.

An additional feature of the present experiments was that the test phase was divided into two subphases. In the first test phase, the participants were told to base their decisions on what seemed more familiar or better. Immediately after this test, they were asked what criteria they used for making their decisions. The aim was to assess their awareness at the level of understanding. In the second test phase, they were told that there was a rule that determined which of the alternatives was correct. They then repeated the test items from the first test phase, in a somewhat different order, but this time with a view to working out what this rule was. Again, no feedback was provided. The purpose of this second test phase was to see if participants who did well on generalization items in the first test phase were more likely to become aware of the rules when instructed to search for them in the second test phase. This would provide evidence that implicit learning can facilitate subsequent explicit learning. Mathews et al. (1989) found that participants who first performed an implicit training task were more likely to become aware of the rule system in a subsequent rule discovery task than those who did not, and they suggested that there might be a synergistic relationship between implicit and explicit learning processes. After this second test phase, the participants were again asked what criteria they used for making their decisions. If they did not mention animacy, the system was explained to them and they were asked whether they had become aware of the potential relevance of animacy at any point during the previous training or test tasks. Although the use of retrospective self-report is clearly not ideal (see Leow, 1997, for discussion), given the simplicity of the present system and the ease of conceptualizing the rules, it was assumed that participants would be able to recall whether any relevant hypotheses had occurred to them during the preceding tasks.

EXPERIMENT 1

Method

Materials.

The noun phrases used for training and testing are shown in Table 1. The system is similar to that used in Williams (2004, Experiment 2), but in this version, there were twice as many nouns per class. The four novel words gi, ro, ul, and ne were selected so as to be unlike any determiners in languages that the participants were likely to know. They were assigned animacy values in such a way as to minimize confounds with sound. The words containing front vowels (gi and ne) expressed opposite values of animacy and distance, as did the two words containing back vowels (ro and ul). If, for instance, gi and ne had both been assigned to the animate category and ro and ul to the inanimate category, then animacy would have been confounded with vowel type. As it was, any such biases worked in the opposite direction to animacy.

Items used in Experiment 1

Nouns were selected so as to be clear cases of animate or inanimate objects. Natural things that do not typically move (e.g., trees or rocks) and man-made things that do move (e.g., bicycles or cars) were not used. What is referred to here as animacy could also be described as living versus nonliving, natural versus man-made, or moves versus does-not-move. There were 12 nouns of each class. Eight nouns of each class occurred with both possible determiners during training but in either singular or plural form (see Table 1). The remaining four nouns of each class only occurred with one determiner during training. These will be referred to as critical items, and the withheld combinations will be referred to as generalization test items. So that critical items occurred as often as the other items during training, they occurred in singular and plural forms (see Table 1). Across the training set, singular and plural nouns appeared equally often with each type of determiner.

Sentence contexts were written for each noun phrase from the training set. The sentences were intended to make it clear how far away the object was in the context of the situation described. A complete listing of the training sentences appears in Appendix A. The training items were divided into two sets, referred to as Set 1 and Set 2. Each noun appeared once in each set, but with either a different determiner or, for the critical items, with a different value of singular or plural. Each set contained all the nouns from each class, including all four critical items.

The materials for the test phase comprised the singular and plural forms of the withheld items from the training phase plus an additional 10 items (5 from each class) that had been presented during training (see Table 1). These will be referred to as generalization and trained test items, respectively. Sentence contexts were written for the test items; these sentences contained verbs that did not occur in the training items. The target noun phrase occurred at the end of the sentence.

After testing a number of participants, it was decided that it would be more prudent to ensure that the general situations described by the test and training sentences were as dissimilar as possible. For example, two of the original test items used the verbs repair and fix, and the correct response was ro. In the training set, ro had occurred once in the context of restore and once in the context of broke. It was decided to revise some of the test items so that the events they described were totally dissimilar to any events in the training set. These will be referred to as the revised test items. Some changes were also made to other items to strengthen the context. All of the test items are presented in Appendix B.

A second version of the training set was constructed, in which the assignment of the determiners to nouns was changed so that gi and ul occurred with inanimates and ro and ne occurred with animates. The same sentence contexts were used. This was to control for the possibility that certain sound combinations might be preferred when participants are forced to choose between alternatives in the test phase. For example, participants might prefer ne lion over ul lion simply because they do not like the combination of an l-final determiner and an l-initial noun. The test items were identical to those in the first version. Half of the participants were tested on each version.

Procedure.

Language background questionnaire. The participants first filled out a questionnaire concerning their age, field of study, L1, and knowledge of L2s. For L2s, the questionnaire elicited information concerning instructional context, qualifications received, contexts of use, and a self-assessment of proficiency (beginner, intermediate, upper intermediate, or advanced).

Phonological short-term memory test. The same PSTM test was used as in Williams and Lovatt (2003) and Williams (2004). This was an immediate serial recall task in which lists of three nonsense words were presented (both auditorily and visually) and the participant had to recall the nonsense words in each list in the order of presentation. The lists were composed from a pool of eight nonsense words: johombe, zabide, wakime, migene, shosane, tisseke, chakume, and nawase. There were 10 practice lists using different nonsense words (including the determiners used in Williams, 2004), followed by 13 experimental lists. One mark was awarded for each word recalled correctly and in the correct list position; half a mark was awarded for a single phoneme error or for a position error.

Vocabulary pretraining. Participants were presented with four cards that had a novel word printed on one side (e.g., gi) and an English word printed on the other (e.g., near). To distinguish the alternative forms for each meaning, the English words were printed in different colors (red, blue, gray, and green for gi, ro, ul, and ne, respectively). The participants studied the cards until they were able produce the novel word when given an English word in a certain color (e.g., to produce gi when presented with near in red). They then practiced this task in a computerized version until they could perform without error. Following this, the participants performed a comprehension task on the items. The computer played a recording of the author saying one of the words (e.g., gi), and the participants had to indicate whether it meant “near” or “far” by pressing one of two keys on the keyboard. A beep signaled an error. It was pointed out to them that the color scheme used in the previous task was now irrelevant, as it was only necessary to classify the words as having been paired with “near” or “far.” There were 12 repetitions of each item presented in a random order.

Training task. The participants were provided with written instructions that explained the general purpose of the experiment as follows:

This is an experiment about how people formulate sentences in different languages. Different languages have different ways of saying things, which raises the question of whether this forces people to think slightly differently in order to speak and understand in different languages. Here I am investigating this issue in a situation where the sentences are almost entirely in English, apart from the four words that you have just learned.

The function of the novel words was then described in the context of the example sentence The little boy patted gi tiger in the zoo (note that this is not one of the training items):

In this language, each time an object is mentioned, it is necessary to specify how far away it is from the subject of the sentence, whether it is “near” or “far.” The word that is used to do this also functions like the English word “the.” So, saying gi tiger is like saying “the-near tiger.” In the context of this example, the word expresses the idea that the boy is near to the tiger (obviously, because he is patting it).

They were then told that their tasks were (a) to listen to each sentence, and on hearing one of the novel words, to indicate whether it means “near” or “far” by pressing the appropriate key; (b) after each sentence, to repeat it exactly aloud; and (c) during (a) and (b), to form a mental image of the general situation described by the sentence, paying particular attention to the distance relationship expressed by the novel word. They were told how many different sentences there would be, with three repetitions of each, and that there would be a memory test for some of the sentences at the end. However, it was stressed that it was not necessary to remember the exact wording of the sentences; it would be sufficient to understand the general situation described by the sentences, which was why it was important to form a mental image as they listened to them. These instructions were designed to deter the participants from trying to remember the precise forms of the novel words and to concentrate instead on their meanings and how they contributed to their understanding of the overall situation described by the sentence. In the course of running the experiment, it was found that some of the nonnatives found it arduous to repeat the sentences in their entirety. In such cases, they were instructed to repeat only the relevant noun phrase (e.g., gi lion) while still forming an image of the overall situation.

The order of trials within each set was independently randomized for each participant. The training phase consisted of six blocks of 24 trials in the order Set 1, Set 2, Set 1, Set 2, Set 1, Set 2. There was a rest break for a minimum of 30 seconds between each block. Because of the distribution of items over sets, the same noun never occurred in consecutive trials, thereby reducing the possibility of making explicit comparisons between items.

Testing phase. After the training phase, the participants were told that before performing the memory test, they would have to perform an exercise on new sentences. For each test item, the computer first displayed the sentence context (e.g., The lady spent many hours sewing …). Upon clicking a continue button, two alternative completions were presented, one above the other (gi cushions / ro cushions). The participants were told to select the noun phrase that seemed “more familiar, better, or more appropriate” on the basis of what they had heard during the training task. The correct phrase occurred as the upper or lower choice with equal frequency.

The testing items were presented in the same order for all participants. This order was contrived in order to reduce the possibility of making comparisons between successive items (e.g., a “far” animate followed by a “far” inanimate). The first test began with two trained items, followed by eight generalization items. Each generalization noun occurred once, and there were equal numbers of singular and plural nouns with each determiner. There were then eight more trained items, followed again by the eight generalization items, but this time, the generalization items showed the opposite value of singular or plural and were in different sentence contexts. Appendix B lists the test items in the order of presentation in this test.

After completing the first test phase, the participants were asked what criteria they had used to make their choices. Any references to living or nonliving, moves or does-not-move, and so forth were interpreted as evidence of awareness of animacy. In such cases, the participants were asked at what point in the experiment they had become aware that animacy was relevant. Participants who did not mention animacy were then told that in fact there was a rule that determined the correct choice and that they were to go through the test items again, but this time, with a view to working out what the rule was. It was explained that this time, the test would begin with items that should seem more familiar and that they should therefore be able to use their responses to these items as a basis for forming hypotheses. This second test phase began with 10 trained items, followed by the first set of generalization items, and then the second set of generalization items. The same sentence contexts were used as in the first test phase. Afterward, the participants were asked for their ideas about what the rule might be; if animacy was mentioned, they were then asked at what point in the experiment they had become aware of its relevance. If they did not mention animacy, the system was explained to them, and they were asked if at any point during the experiment they had considered the possible relevance of animacy.

Memory test. A yes/no sentence recognition test was used as the memory test. There were 10 sentences from the training set and 10 new sentences. The items were presented by computer and in a different random order for each participant.

The entire procedure took about 1 hour.

Results

A total of 8 of the 41 participants expressed awareness of the relevance of animacy when interviewed at the end of the first test phase. They were also able to state the specific relationships between the determiners and noun animacy. Six of them had become aware of the rule during the training phase, and two of them had only worked out the rule during the first testing phase.²

It was noted that of the six participants who became aware during the training phase, five were native speakers of English. This possibly reflects the lower processing demands of performing the training task in the native language, making it more likely that the rule will be noticed.

As one would expect, their generalization test performance was almost perfect (91.4% averaged over the two sets of generalization items in the first test phase, and significantly above chance, p < .001).

Of the remaining 33 participants, 6 described some kind of rule, but with low levels of confidence. None of them mentioned animacy, but rather referred to the syntactic structure of the sentence, the phonology of the noun, the nature of the action (without being more specific), or the directness of contact involved in the action (and, in one case, also the visibility of the object in the case of “far” items). The remaining 27 claimed to have based their decisions on familiarity or intuition. When told what the rule was, all 33 participants claimed not to have considered the relevance of animacy at any point during the experiment. They were, therefore, classified as unaware with respect to the target rule (i.e., as having no awareness at the level of understanding). Table 2 shows their mean scores in the first test phase, grouped by the components of the test in the order that they occurred in the experiment (with the exception of the first two trained items; these have been combined with trained 1). The difference between scores in each condition and chance (50%) was evaluated using single group t-tests. The mean over the two sets of generalization items in the first test phase is also shown (mean generalization). All test scores were significantly above chance, even for generalization items, where neither alternative had, in fact, been presented during training.³

Despite showing above-chance performance, the unaware group performed significantly below the level of the aware group on all tests in the first test phase: generalization 1 (61.7% vs. 84.4%, p < .05), trained items (70.4% vs. 90.0%, p < .01), generalization 2 (59.8% vs. 98.5%, p < .001).

Mean percentage correct, standard deviations, and t-values for deviation from chance (t): 33 unaware participants (Experiment 1, first test phase)

Interestingly, the six participants who reported non-animacy-based rules performed poorly, with a mean generalization test score of 49%. In contrast, they scored 68% on trained items, which is significantly above chance, t(5) = 3.84, p < .05, and similar to the level of performance of the overall sample. Thus, the generalization test scores in Table 2 are actually being depressed by the inclusion of the participants who reported erroneous rules. If they are excluded, the generalization scores become: generalization 1, 65%, p < .001; generalization 2, 62%, p < .01; mean generalization, 63%, p < .001.

Recall that the original set of generalization test items contained a few verbs that were semantically similar to verbs that had occurred during training, whereas in the revised test, all of the generalization items used verbs that were dissimilar to those used in training. There was no evidence that this change in the materials influenced the results. For the 18 of the 33 unaware participants who received the original test, the mean generalization test score was 62.1%, which was significantly above chance, p < .01. For the 15 participants who received the revised test, the mean generalization score was 59.2%, which was again significantly above chance, p < .05.⁴

The participants who received the revised version knew slightly fewer gender languages than those in the original version (1.53 and 1.89, respectively). Given the positive correlation between knowledge of gender languages and learning (see discussion), this might explain the slightly lower level of learning for participants who received the revised version.

Also, there was no effect of the language version that the participants received. For example, for the 16 participants who received the version in which gi and ul were animate, the mean generalization performance was 59% (significantly above chance, p < .05), and for the 17 participants who received the version in which ro and ne were animate, the mean generalization performance was 62.5% (significantly above chance, p < .01).

It is possible that the above-chance level of responding to generalization items was just due to a few items where the sentence context perhaps reminded the participants of a relevant training sentence. On the other hand, if the effect represents an animacy bias, then it should be present for all of the test items. Therefore, test performance was analyzed by items and was found to be significantly above chance for the 16 generalization items, p < 0.001, and for the 10 trained items, p < .001. Of the 16 generalization items, 13 had mean scores above chance, 2 were at 50%, and 1 was less than 50%. Clearly then, the effects were not due to just a few items.

In the second test phase, the unaware participants were instructed to search for a rule that determined the use of the novel words. Following this exercise, a total of 11 participants were aware of the system, and the remaining 22 were either unable to report a rule or suggested other rules with low levels of confidence. The scores for these two groups over both test phases are compared in Table 3.

Performance of the participants who were unaware of the rule after the first test phase of Experiment 1 according to whether they expressed awareness after the second test phase

As would be expected, the participants who were aware of the system at the end of this second test phase had significantly higher scores for the trained (trained 2) and generalization (generalization 3) items than those who were not aware, as confirmed by independent sample t-tests (see the bottom row of Table 3). However, the performance for the 22 participants who were still unaware of the rule at the end of this phase was still significantly above chance for both trained items and generalization items.

It is now interesting to look back at the generalization test performance for these two groups in the first test phase. Did the participants who became aware of the rule in the second test phase show stronger evidence for implicit learning in the first test phase? There was some evidence of this for the first set of generalization items, but the difference between the aware and unaware groups does not quite reach significance, p = .058. Single group t-tests showed that at generalization 1, only the score for the aware group was significantly above chance. There is some evidence here, therefore, that those participants who became aware in the second test phase showed more evidence for implicit learning in the first test phase—at least when assessed by the first set of generalization items. Note, however, that even participants who remained unaware at the end of the second test phase showed above-chance generalization test performance in the first test phase, at least for the second set of items, and when the mean of the two sets of generalization items is considered..

For the overall sample of participants, the mean score on the PSTM test was 70%, and the mean number of gender languages known to an intermediate level or better was 1.68 (SD = 1.29). The analysis of individual differences first focused on the mean generalization performance in the first test phase for the 33 participants who claimed to be unaware at that point. The performance of the 11 participants who spoke a gender L1 was not significantly different from the 22 participants who did not [64% vs. 59%, t(31) = 0.86]. However, performance for the 12 participants studying language-related disciplines—the linguists—was significantly better than for the 21 nonlinguists [69.3% vs. 55.9%, t(31) = 2.47, p < .05]. Next, correlations between mean phase 1 generalization performance and the following factors were calculated: performance on the PSTM test, r = .01, the number of L2s spoken to an intermediate level or better, r = .263, and the number of gender languages spoken to an intermediate level or better, including L1, r = 0.357, p < .05. Therefore, both studying linguistics and knowing gender languages—but not necessarily as an L1—were associated with better performance. However, these two factors are also somewhat related given that the linguists knew more gender languages than the nonlinguists (a mean of 2.17 as opposed to 1.48), although the difference was not significant, t(31) = 1.55.

EXPERIMENT 2

Method

Materials and Design.

The noun phrases used in the experiment are shown in Table 4. The critical items, generalization items, and trained test items were all the same as in Experiment 1 so that exactly the same test material could be used. The only difference was that during training, each noun only occurred with one determiner. So that there would be the same number of different sentences as in Experiment 1, each noun occurred in both singular and plural forms and in a different sentence context in each case. The materials were divided into two sets, as for Experiment 1, so that each noun occurred once in each set and with half singular and plural forms in each set. The items and sentence contexts in the first set were the same as those used in Experiment 1. New sentence contexts were written for the items in the second set (see Appendix A).

Items used in Experiment 2

A second version of the training set was constructed in which the assignment of the determiners to nouns was changed so that gi and ul occurred with inanimates and ro and ne occurred with animates. Equal numbers of participants received each version. The testing material was exactly the same as the revised version of Experiment 1. The test sentences contained different verbs and described different situations from the training sentences.

Procedure.

The procedure was identical to Experiment 1. The experiment again lasted about 1 hour.

Results

Seven of the 24 participants expressed awareness of the relevance of animacy and were able to describe the rules when interviewed at the end of the first test phase. Two of them had become aware during the training phase, and the remaining five reported that the rule “just came to them” during the first test phase. The mean first test phase generalization performance for this aware group was 85%, significantly above chance, p < .01.⁶

One participant in this group expressed awareness of the relevance of animacy but reported the wrong mapping onto the words for “far.” Her mean performance on the generalization items was 44%. The mean over the remaining six aware participants was 92%.

It is interesting that for the five participants who became aware during the test phase, performance over the first set of generalization items, constituting the first third of the test, was 75%, significantly above chance, p < .05. Either rule awareness developed very quickly or else accurate responding by intuition led to subsequent rule awareness.

The scores for the remaining 17 unaware participants in the first test phase are shown in Table 5. Significant differences from the chance level of performance are also indicated. All scores were well above chance, even for the generalization items.⁷

Scores for the unaware group were significantly worse than the aware group, although only for the generalization items: generalization 1 (64.0% vs. 82.1%, p < .01), trained items (84.7% vs. 91.4%, p = .20), generalization 2 (64.0% vs. 87.5%, p < .01).

Items analyses were also performed. The mean score over the 16 generalization test items was significantly above chance, p < .001, as was the score over the trained items, p < .001. Of the 16 generalization items, 13 had above-chance scores, 1 was slightly below chance, and 2 were at chance. Only one of these items was also at chance in Experiment 1.

Mean percentage correct, standard deviations, and t-values for deviation from chance (t): 17 unaware participants (Experiment 2: first test phase)

In the second test phase, only three of the previously unaware participants were able to correctly state the rule. This made it impossible to make meaningful comparisons in performance over the first test phase. The situation was also complicated by the fact that a further three participants were classed as partially aware because they had realized the relevance of animacy but only stated the correct relationship to animacy for one pair of determiners (i.e., either the “near” or “far” pair). Their performance was poor over all three sets of generalization test items (54%, 46%, and 48%). What was clear, however, was that phase 1 generalization test performance was well above chance for the remaining 11 participants who still claimed to be unaware after the end of the second test phase. The data are shown in Table 6.

Mean percentage correct, standard deviations, and t-values for deviation from chance (t): 11 unaware participants (after Experiment 2, second test phase)

For the whole sample, the mean score on the PSTM test was 71% and the mean number of gender languages known to an intermediate level or better was 1.42 (SD = 1.47).

The analysis of individual differences focused on the mean phase 1 generalization test performance for the 17 participants who claimed to be unaware at that point. This time, the performance of the eight participants who spoke a gender L1 was significantly better than the nine participants who did not, 72% versus 57%, t(15) = 2.48, p < 0.05, although participants who spoke a gender L1 also knew more gender languages overall than those who did not—a mean of 2.0 versus 0.11, t(15) = 5.05, p < .001. This time, performance for the 10 linguists was not significantly better than for the 7 nonlinguists, 66% versus 61%, t(15) = 0.78. Correlations between mean phase 1 generalization performance and the following factors were calculated: PSTM, r = .228, the number of L2s spoken to an intermediate level or better, r = .404, and the number of gender languages spoken to an intermediate level or better, including L1, r = .358. However, it must be recognized that these correlational analyses are highly unreliable because of the small sample size and the limited variation in the number of gender languages known by the unaware participants (M = 1.0, SD = 1.22). Nevertheless, the fact that participants who spoke a gender L1 outperformed those who did not does suggest that—just as in Experiment 1—knowledge of gender languages had some effect.