This paper does a fine job of advancing discussion concerning a question that is indeed quite underrepresented in the literature, that is, how language learners comprehend and produce language in real time. The paper is firmly rooted in the dual mechanism approach to language processing and takes as its starting point the assumption that normal adult processing is characterized by two systems, one that is lexically based and one that is essentially combinatorial. The authors cite evidence that both first language (L1) learners and adult native speakers show evidence of dual mechanism processing and that, in particular, children's sentence processing shows early reliance on structure-based interpretation and less ability to employ lexical/pragmatic information in the resolution of language ambiguity. One way to view this preference is that L1 learners might know, broadly speaking, considerably more about their language than they do about the world in which they live. Adult second language (L2) learners might be said to be in exactly the opposite situation. It is therefore hardly surprising that adult L2 speakers rely strongly on lexical/pragmatic cues in sentence processing. In the early stages of adult L2 acquisition, the demands of real-time processing make use of such nonsyntactic inference crucial. The question that strikes me as key is whether, as L2 speakers become more proficient, they are weaned from this reliance such that their processing reflects the interaction between syntactic and lexical processing that is characteristic of adult native speakers. When and if they do, we could say that their L2 processing is, both internally and externally, nativelike.

We might represent this characterization in simplified form as follows:

We know, by definition, that Example 1a becomes Example 1c, and Clahsen and Felser (CF) argue that this transition involves quantitative rather than qualitative changes along the way within fundamentally the same dual mechanism architecture. The mystery at hand is why, even at very high levels of L2 proficiency, Example 1b does not become Example 1c.

CF present two possible reasons for this: the first is the Paradis/Ullman perspective (Paradis, 1994, 1997, 2004; Ullman, 2001), which states that L1 language is characterized by the dominance of procedural knowledge and that L2 language is characterized by the dominance of declarative knowledge. The second is the shallow structure hypothesis (SSH), which they propose.

Although I am sympathetic to the authors' view that characterizing L1/L2 differences solely in terms of procedural versus declarative knowledge may be too simplistic, I am not entirely convinced that their arguments warrant its dismissal. The model is indeed underspecified, but it must be admitted that their alternative account is, at present, also a sketch. They argue that the Paradis/Ullman (Paradis, 1994, 1997, 2004; Ullman, 2001) position could not account for the observed distinction between L2 learners' morphological and syntactic abilities, and that we are left wondering what exactly concepts such as “reduced availability of procedural knowledge” might mean. It seems to me at least possible that, in this context, “reduced availability” might mean exactly that procedural knowledge is insufficiently available to deal with the computational complexity posed by sentences as compared to relatively simple inflected forms.

Even if the arguments for rejecting the Paradis/Ullman (Paradis, 1994, 1997, 2004; Ullman, 2001) position are less than overwhelming, the authors' alternative SSH does seem to have substantial merit. The hypothesis provides a very plausible account for why Example 1b does not become Example 1c, namely, because the requisite structural complexity required to scaffold syntactic processing is not available to the L2 learner. In this way, their hypothesis is reminiscent of Caplan and Hildebrandt's (1988) account of sentence comprehension disorders in aphasia, positing a flat rather than hierarchical syntactic representation upon which sentence interpretation can be based.

This may be simply another way of saying that procedural knowledge is less available. However, even if it is, I find it to be one that is more useful in that it casts the phenomenon in terms that are likely to lead to further refinement of our understanding. I am struck, for example, by the possibility that the authors have set up a false dichotomy between morphological and sentence processing. As the authors note, the observed differences between morphological processing and sentence processing may be related to the fact that the morphological structures studied so far in this domain “involve the simple concatenation of two adjacent elements (stems and affixes) and thus have a much shallower internal structure than sentences.” Perhaps, then, the differences that the authors note between morphological and sentence processing are not differences that are intrinsic to these domains, but simply ones that fall out as artifacts of the stimulus materials chosen among studies. Indeed, it seems to me that German, the language in which the authors have done so much work, offers fertile ground for investigating the interaction of lexical/pragmatic knowledge and hierarchical syntactic knowledge within the morphological domain. Unlike English, German allows for hierarchically structured triconstituent compounds to be written as single words, so that a word such as Lederhandschuh (leather + hand + shoe [leather glove]) clearly has a right-branching structure, whereas Handschuhleder (hand + shoe + leather [glove leather]) clearly has a left-branching structure.

Krott et al. (2004) have presented evidence that, all things being equal, native speakers of German show a preference for left-branching analyses of triconstituent compounds. Against this background, lexical/pragmatic constraints could be manipulated to yield contrasts in the morphological domain that might be comparable to those that have been used to investigate the interaction of structural and lexical/pragmatic preferences in the sentence domain.

As an example of this, consider the four novel compound structures in Example 2. The compound in Example 2a is ambiguous in that it could plausibly have both a left- and right-branching structure. Structure 2b is also balanced, but in a manner such that neither reading corresponds to our real-world knowledge of the materials typically used to manufacture either carpets or brooms. Finally, in Structures 2c and 2d, the choice of the initial constituent pragmatically biases toward either a left-branching or right-branching reading.

My expectation is that stimuli such as these in the morphological domain would yield very much the same pattern of data that has been reported in sentence processing for adult native speakers of German, L1 learners, and adult L2 learners. L1 learners should show a left-branching (structural) bias, which is relatively insensitive to pragmatic constraints, whereas L2 learners should show a reliance on pragmatic cues and less of a structurally based left-branching preference. If this prediction is correct, it would suggest no underlying morphological/sentence processing difference and the need for a unified account across the domains.

Finally, I would like to note that I was very pleased by the authors' comment at the end of the paper concerning the need to investigate L2 processing in real time across different levels of L2 proficiency. This will be a critical step in the development of our understanding of whether structural depth follows a developmental pattern in L2, or whether, as the authors seem to suggest, shallow structure is tied to the fundamental nature of L2 processing.