Understanding the mechanisms learners use to process target language input is crucial to developing a complete model of both first language (L1) and second language (L2) acquisition. If adult L2 learners are found to process the target language with mechanisms that differ from those used by child L1 learners and adult native speakers, what implications might this have for the developing grammar? Clahsen and Felser review evidence that appears to point to such differences, generalizing their findings under a shallow structure hypothesis about how adult learners process input in L2.

The claim is that adult learners project a “shallower” structure than do child L1 learners or adult native speakers. Such structure contains less detail about certain types of dependencies (e.g., filler-gap chains), and may not fully resolve certain types of ambiguities (e.g., relative clause attachments), although adult L2 learners can nevertheless exploit nonstructural information to determine meaning. The shallow structure hypothesis thus seems to claim that a person who has learned an L2 in adulthood (after the onset of puberty, presumably) uses one type of processing mechanism for input in L1, a different type of mechanism for input in L2. (I have described this as language-dependent processing; Fernández, 1998, 2003.)

Let us consider the issue of L2 processing within a larger theoretical framework. In a perfect world, L1 and L2 acquisition would be fundamentally similar (see VanPatten, 2005). Likewise, any type of linguistic input, in L1 or L2, would be processed in fundamentally similar ways, with one mechanism that follows the same operating principles. These are traditionally described as principles of minimal effort, motivated by the inherent limitations of human cognition (e.g., Frazier, 1978). However, the world may not be all that perfect; the human language processor might be found to have different operational properties, depending on whether the stimulus language is an L1 learned in childhood or an L2 learned in adulthood, as the shallow structure hypothesis seems to claim. If so, we face having to explain not only how and when the processing mechanism arrives at such a change as the individual matures, but also what consequences such a change will have for the acquisition of the target language grammar. If the shallow structure hypothesis makes accurate predictions about adult behavior with L2 input, shallow performance could be one source for the difficulties adults encounter when learning an L2 (e.g., Johnson & Newport, 1989), an interesting possibility that has not accrued much empirical support (but that has certainly been contemplated before; see, e.g., Dussias, 2003; Fernández, 1999).

We now turn to one phenomenon Clahsen and Felser investigate to examine L2 processing: the relative clause (RC) attachment ambiguity. In this construction, a restrictive RC may modify one of two nouns in the complex noun phrase (NP) that precedes it, for example:

1. The dean liked the secretary of the professor who was reading a letter.

The RC attachment ambiguity has been the focus of many studies of L2 processing for at least one important reason: monolingual speakers of different languages have been found to differ in how they prefer to interpret the ambiguous string, and such cross-linguistic differences permit asking whether the system used to process L1 might differ from the system used to process L2.

Since the first report of the phenomenon (Cuetos & Mitchell, 1988), many studies have shown that monolingual speakers of some languages (English, most notably) prefer to interpret sentences like Example 1 with the RC modifying the second noun, actress (a “low attachment” interpretation), while monolingual speakers of other languages (German, Greek, Russian, Spanish, to name a few) prefer for the RC to modify the first noun, maid (a “high attachment” interpretation). The mechanism responsible for building syntactic structure, the parser, likes to build minimal trees (e.g., Frazier, 1978): it prefers attachments that involve fewer nodes (minimal attachment), that create shorter filler-gap chains (minimal chains), and that are made locally (late closure). (Minimal structure should not be confused with the notion of shallow, “less detailed” structure developed by Clahsen and Felser.) The simpler tree for Example 1 would be one where the RC is attached low to the local noun, actress, following the principle of late closure (Frazier, 1978). Attaching to the higher and farther noun, maid, means violating this preference for local attachments. Why would speakers of some languages exhibit such behavior?

Of the many proposals designed to answer this question, those with the most solid empirical support assume that the parser operates the same way in all languages. Such proposals claim that some languages prefer the (theoretically) more difficult high attachments because of extrasyntactic factors. The catalyst promoting high attachment could be a grammatical property of the language, for example, prosody (Fodor, 2002), or the language-specific consequences of the application of discourse principles (Frazier & Clifton, 1996). Details about their differences aside, these two accounts (prosody, discourse) agree in that the cross-linguistic differences are not sourced in first-pass syntax-only structure-building processes, but rather have to do with considerations outside of the parser proper. In support of this idea, robust cross-linguistic differences have been reported in studies employing untimed, or “off-line,” methods (see, e.g., Fernández, 2003, and references there), methods that allow for such extrasyntactic factors to be taken into consideration. Indeed, “on-line” studies have yielded a rather mixed set of findings (some of the controversies are discussed in Fernández, 2003), plausibly because there is no guarantee that a given “on-line” method effectively prevents the system from considering extrasyntactic information. (The RC attachment ambiguity may therefore not be of theoretical or practical interest, if the goal is to examine early processing routines in L2. Nonetheless, studying how final interpretations are arrived at might be of use for developing an understanding how different types of learners use different types of information to determine meaning, regardless of how early or late this information might be used.)

If the source of the cross-linguistic differences lies outside of the parser, possible outcomes, in studies of RC attachment in L2, are restricted to those where a learner will have similar preferences in L2 and L1, unless knowledge of the critical component (prosody, discourse) is nativelike in both languages. (Note: when L1 coexists with an L2, there is no a priori reason to assume the system remains intact and is not affected either by L2's grammar or by the cognitive load of having two grammars and lexicons in place. That an L1 speaker of a high attaching language will continue attaching high in L1, regardless of attachment preference for stimuli in L2, is an assumption that requires empirical confirmation.) Thus, we expect to find language-independent processing (Fernández, 1998, 2003), absent nativelike knowledge in L2 of whatever property guides the language's attachment preferences. Crucially, we need to compare directly L2 and L1 RC attachment preferences.

The RC attachment data offered by Clahsen and Felser in support of the shallow structure hypothesis come from separate experiments carried out in low-attaching English (Felser, Roberts, Gross, & Marinis, 2003) and high-attaching Greek (Papadopoulou & Clahsen, 2003). In both studies, the behavior of native speakers is contrasted to that of learners whose L1 is a language that either differs in preference from L2 (German or Greek L1 speakers learning English; Felser et al. 2003) or not (German, Russian, and Spanish L2 speakers learning Greek; Papadopoulou & Clahsen, 2003). All experiments tested target materials like Example 1, above, and Example 2:

1. The dean liked the professor with the secretary who was reading a letter.

The contrast between Examples 1 and 2 has to do with the nature of the preposition in the complex NP: nonthematic of in Example 1 (and its equivalent in the genitive construction used in the Greek studies) versus thematic with in Example 2 (and its equivalent in Greek). The preposition in Example 2 is more likely to block high attachment than the preposition in Example 1.¹

The general findings of these experiments are straightforward (see also Clahsen & Felser's table 4). In both untimed (off-line) questionnaires and grammaticality judgment tasks, and timed (on-line) self-paced reading tasks, Clahsen, Felser, and colleagues find that speakers of all language history profiles attach low with materials like Example 2, all being similarly influenced by the presence of a preposition like with, whatever the source for this may be (see note 1). The interesting findings come from responses to materials like Example 1, for which these experiments replicate the standard cross-linguistic difference with monolinguals (even in the on-line tasks): Greek monolinguals attach high, English monolinguals attach low. However, with sentences like Example 1, the L2 learners, across the board, apparently fail to exhibit a preference for either site, regardless of native language, and regardless of stimulus language. Where we would expect similarity between how we assume these learners process their L1s and how they process their L2s, we find instead that L1 speakers of high attachment languages fail to show a preference to attach high in L2, whether L2 is high-attaching Greek (Papadopoulou & Clahsen, 2003), or low-attaching English (Felser et al., 2003).

To support the shallow structure hypothesis, we are forced to stipulate that the L2 learners would perform, in the same task and with fully comparable materials, in their respective L1s as do the monolingual controls. Such a stipulation becomes problematic in light of evidence from a number of studies that have compared directly how speakers of two languages process ambiguous RCs in both languages. These have systematically found similarity in attachment preference within speakers across languages (Dussias, 2003; Fernández, 2003; Frenck-Mestre, 2002; Maia & Maia, 2005).

Clahsen and Felser argue that methodology is not to blame for the apparent discrepancies in the empirical database, pointing to the similar outcomes of their on-line and off-line tasks. Taking into consideration that cross-linguistic differences in monolinguals are only robust when the task is off-line, we could very well be suspect of Clahsen and Felser's on-line data, but we should accept their off-line data. Yet, the L2 Greek off-line data come from a task that differs considerably from those used elsewhere: it requires participants to rate the acceptability of sentences where number agreement forces the RC to attach high or low. We have no way of telling if L2 learners make acceptability judgments with the same confidence as do native speakers (see discussion and references in Altenberg & Vago, 2004), or if L2 learners use agreement features such as number with the same efficacy as do native speakers. Thus, the only data we can safely evaluate against existing studies comes from the English questionnaire reported by Felser et al. (2003), where participants read fully ambiguous stimuli and selected one of two possible interpretations for the RC (as in other studies). Here, the only statistical tests performed on the quantified data, number of N2 responses, were one-sample t tests, which show that for materials like Example 2 performance was above chance, for all three groups (L1 English, German, and Greek), while for materials like Example 1 performance was above chance for the native speakers, at chance for the two groups of L2 learners. A more appropriate treatment for these data (somewhat problematic, given the uneven number of participants, N = 45 for L1 English, N = 28 for L1 German, N = 39 for L1 Greek) would be a two-factor analysis of variance (Speaker Group × Preposition). Given the means reported, we would expect reliable main effects of both speaker groups (N2 responses average at 77% for L1 English, 70% for L1 German, 67% for L1 Greek) and preposition (55% for nonthematic of, 87% for thematic with). The shallow structure hypothesis would predict an interaction of the two factors, with similar preferences for all groups for materials like Example 2, but differences between learners and native speakers for materials like Example 1. The means reported do not suggest the presence of such an interaction: overall, the native speakers chose N2 interpretations more frequently (9%, on average), in both conditions, an effect that remains without explanation. Notice also that the analyses performed by Felser et al. (2003) make the assumption that chance (50% N2 responses) equals no preference. With data collected using unnormed materials, this is by no means guaranteed: a given materials set might be intrinsically biased toward one or the other attachment site (Scheepers, 2003).

Clahsen and Felser point out that their experiments have an important mechanism in place to ensure that the L2 learners had knowledge of the target language grammar (they were administered a general proficiency test) and of the grammatical principles that generate RCs in L2 (they were also administered a grammaticality judgment task probing relevant constructions). Yet, these tests may not be sensitive to the language-specific factor responsible for the cross-linguistic differences of interest here. Two factors to consider are prosody and discourse principles, neither of which is typically tapped by general proficiency tests or by grammaticality judgment tasks. The reported “random” behavior of the learners might have been guided by the misapplication by some, and adequate application by others, of L2 nativelike prosodic or pragmatic principles. It is not at all far-fetched to imagine that these components of language (prosody, discourse) are not fully learned until very late in the acquisition process, and are subject to high between-speaker variability. How to test for L2 knowledge of such elusive components is unclear; what is clear is that we urgently need to develop such tests, if the goal is to understand the source for the behavior of L2 learners in experiments such as these.

The L2 RC attachment data offered by Clahsen and Felser thus fail to convince me that L2 processing differs fundamentally from L1 processing. At the same time, I acknowledge the powerful implications of the shallow structure hypothesis, in its claim that the mechanism for processing input in a language being learned changes as the individual matures. We would need to determine how, and when, this change takes place, and what effect such a change would have for the acquisition of the target language. Shallow performance might turn out to be the source for the variable success rates in L2 acquisition, under a model in which the development of competence is guided by universal grammar, but in which the shallow processing of input in L2 induces a departure from the natural acquisition process. It is a hypothesis that merits further empirical testing, precisely the type of enterprise that advances science.