Language as skill: Intertwining comprehension and production

Are comprehension and production a single, integrated skill, or are they separate processes drawing on a shared abstract knowledge of language? We argue that a fundamental constraint on memory, the Now-or-Never bottleneck, implies that language processing is incremental and that language learning occurs on-line. These properties are difficult to reconcile with the ‘abstract knowledge’ viewpoint, and crucially suggest that language comprehension and production are facets of a unitary skill. This viewpoint is exemplified in the Chunk-Based Learner, a computational acquisition model that processes incrementally and learns on-line. The model both parses and produces language; and implements the idea that language acquisition is nothing more than learning to process. We suggest that the Now-or-Never bottleneck also provides a strong motivation for unified perception–production models in other domains of communication and cognition. 2016 Published by Elsevier Inc. Language as knowledge; language as skill The ability to comprehend and produce language requires spectacular levels of skill. But is it one skill, or two? Transfer between comprehension and production appears to suggest a unitary system; hearing a language is, it seems, crucially important for speaking that language. After all, children don’t simultaneously learn to understand German while producing Mandarin. Indeed, across diverse theoretical perspectives in the language sciences, there is general agreement that there exists an important overlap between the knowledge and processing operations involved in comprehension and production. But there is considerable disagreement about the nature of this relationship. One viewpoint (e.g., Chomsky, 1965), which has been dominant in many theoretical approaches to language, starts by assuming a strong separation between linguistic competence (i.e., an abstract specification of the speaker/ hearer’s knowledge of the language) and linguistic performance (the processes by which this abstract competence is deployed in language processing). From this standpoint, the overlap between production and comprehension may reside purely in the shared abstract linguistic competence that is being drawn upon by both comprehension and production processes. But the processes of comprehension and production could, in principle, be completely unrelated. We call this the ‘language as knowledge’ view. An opposing viewpoint suggests that no such abstract linguistic competence exists—rather, acquiring language is no more than acquiring the ability to process language. There is no separate representation of the abstract structure of the language (e.g., a grammar) distinct from the mechanisms of language production and comprehension; instead there are simply procedures for language processing (e.g., Kempson, Meyer-Viol, & Gabbay, 2001; O’Grady, 2013). From this point of view, the overlap N. Chater et al. / Journal of Memory and Language 89 (2016) 244–254 245 between comprehension and production requires that the same (or highly overlapping) processes underpin both the comprehension and production of utterances. We call this account the ‘language as skill’ perspective (see also Christiansen & Chater, 2016). We will argue, in the next section, Fundamental memory constraints on skill learning, that basic limitations on memory strongly favor the language as skill perspective, and hence the assumption that the processing operations of comprehension and production are intimately related. This requires viewing language acquisition as a process of online skill learning, where processing operations lay down traces that facilitate further processing—there is no opportunity for inferring abstract general principles of language structure. In light of the limitations we outline, the challenge of building cognitively plausible models of language processing and acquisition is considerable. We take some initial steps toward addressing this challenge in Section ‘A unified model of production and comprehension’, describing a computational model incorporating incremental processing and on-line learning. Crucially, although comprehension and production are closely integrated in the model, we present new simulation results demonstrating that the model gives rise to the kind of comprehen sion–production asymmetry often observed in language acquisition (e.g., Fraser, Bellugi, & Brown, 1963). In Sectio n ‘Integrated production and comprehension’, we then reflect on the broader theoretical issues raised by proposing a unitary model, before drawing brief conclusions. 1 We want to stress that we are not advocating for so-called ‘radical incrementality’ in production, in which words are articulated immediately without any planning ahead. Rather, we see production as involving planning a few chunks ahead at every level of linguistic abstraction. Importantly, though, whereas planning at the level of the phonological word may be quite short in temporal scope, planning will extend further ahead at the level of multiword combinations, and even longer at the conceptual/discourse level (see Chater & Christiansen, in press, for further discussion). Fundamental memory constraints on skill learning On just about any measure, language processing is astonishingly fast. Speaking rates are typically 10–15 phonemes per second, which translates to as much as 150 words per minute (Studdert-Kennedy, 1986). This daunting speed implies that the comprehension system is faced with a relentless onslaught of new input. Similarly, the production system has to generate and execute a stream of articulatory instructions at a remarkable speed. Furthermore, the interleaving of comprehension and production processes is also very fast, as made evident by the rapid turn-taking observed across languages and cultures (e.g., the mean latency between ‘turns’ is typically about 200 ms, Stivers et al., 2009), the ability to ‘shadow’ speech within a 250 ms latency or less (Marslen-Wilson, 1973, 1985), and our ability to fluently complete each other’s sentences (Clark & Wilkes-Gibbs, 1986). Our impressive performance in processing language contrasts strikingly with our very limited ability to process sequences of arbitrary auditory or visual stimuli. For example, in a classic study, Warren, Obusek, Farmer, andWarren (1969) found that naïve listeners were unable correctly to recall the order of just four different non-speech sounds, even though each could easily be identified in isolation. More broadly, memory for the temporal order of arbitrary stimuli appears restricted to 4 ± 1 items (Cowan, 2000); and even the identities of the items in a sequence is typically rapidly lost, e.g., when measured by probed or free recall (e.g., Baddeley, 2007). In short, memory is fleeting: unless information is recoded and/or used rapidly, it is subject to severe interference from an onslaught of new material. We call this the Now-or-Never bottleneck (Christiansen & Chater, in press). To cope with the flow of speech input, it is crucial that phonemes are rapidly recoded into higher-level units—for example, into syllables, lexical items, phrases, and beyond (although the specific hierarchy of linguistic levels is, of course, controversial; and may vary from language to language): we call this Chunk-and-Pass language comprehension (Christiansen & Chater, in press). These more abstract levels correspond both to larger units of speech input (so that the same threeor four-item limit corresponds to a longer stretch of speech), and also will typically lead to less interference with subsequent material. This is because, on just about any theoretical account, the space of lexical items is very much larger than space of phonemes, so that confusability between phonemes will be much greater than confusability between lexical items. Parallel issues arise in speech production. Here, the constraint is to maintain the shortest possible ‘inventory’ of material waiting to be generated, to avoid interference between items to be produced, at a given level of representation (e.g., Dell, Burger, & Svec, 1997). And as with comprehension, this, too, can only be done by decoding higher-level representations into lower-level representations in a piecemeal manner. If, for example, an entire message were decoded into a string of phonemes before even beginning to speak, that stream of phonemes would vastly exceed the few items we can accurately hold in memory, and hence would rapidly be lost. The production system must, therefore, decode a higher-level representation into a more detailed lower-level representation when that lower level representation will be used, and not substantially before. Thus, while our speech production system may look ahead several words in advance, those words will only be converted into, say, a phoneme representation, when the word is almost ready to spoken; and the phonemes will in turn only be translated into still more detailed articulatory instructions at the very last moment. The cascade of different levels of speech production thus obeys a principle of ‘Just in Time’ processing, analogous to the inventory management system pioneered in Japanese manufacturing (Ohno & Mito, 1988). A ‘stock’ of representations cannot be allowed to accumulate, because such stock is highly ‘perishable’—specifically, it will be rapidly interfered with or overwritten by the continual arrival of new material (Christiansen & Chater, in press). The Now-or-Never Bottleneck has particularly striking implications for language acquisition. If linguistic input can only be retained very briefly before being overwritten, then learning must occur as language is being processed. 246 N. Chater et al. / Journal of Memory and Language 89 (2016) 244–254 There is no opportunity for the learner to survey previous linguistic input in order to carry out, for example, distributional tests to determine linguistic categories (e.g., Redington, Chater, & Finch, 1998) or to search for phrase structure rules (Pereira & Schabes, 1992). Indeed, typically, the learner will not be able to take a synoptic view even of a single utterance: the early part of an utterance will typically have been recoded into higher level representations before the end of the utterance is reached. We call the constraint that learning occurs in-the-moment on-line learning (Christiansen & Chater, in press). In this article, we propose that the constraint that language is processed incrementally and learned on-line imposes very strong restrictions on computational models of language processing, and, crucially for the present discussion, suggests an intimate relationship between comprehension and production. Note, in particular, that from the language as skill perspective, it is natural to assume that learning is simply a matter of storing traces of past (incremental) language processing operations (chunks, in the model outlined below). For example, according to instance-based theories of skill learning (e.g., Logan, 1988), lexical access (Goldinger, 1998), and memory (Hintzman, 1986), the processing of new input draws on traces of past processing of previous items. But if this viewpoint is right, the existence of any transfer between comprehension and production must imply that the same processing traces are relevant to both—otherwise, traces from comprehension (e.g., hearing a new word or syntactic construction) could not be recruited in production (e.g., using that same new word or construction). The on-line nature of learning strongly suggests, we argue, a unitary model of comprehension and production. Language processing is a skill; and, to a large extent, comprehension and production are the same skill. Moreover, in this view, language acquisition is nothing more than the learning the skill of language processing. There is, in particular, no additional challenge of acquiring knowledge of the language, over and above the (unitary) ability to comprehend and produce. By contrast, it is difficult to reconcile incremental processing and on-line learning with the ‘‘language as knowledge” viewpoint, according to which comprehension and production correspond potentially to non-overlapping skills drawing on a common body of abstract knowledge of the structure of the whole language (i.e., linguistic competence, Chomsky, 1986). The challenge of extracting a putative abstract linguistic competence most naturally and reliably operates ‘off-line’ by finding regularities over a large corpus of full sentences, in order to find the best overall match between the grammar and corpus (e.g., Klein & Manning, 2004; Pereira, 1992). Yet while the Now-or-Never Bottleneck can be used to provide a theoretical argument for a unitary model of 2 The present paper extends the argument of Christiansen and Chater (in press) to suggest that the Now-or-Never bottleneck most naturally fits with a unified model of comprehension and production, and that language learning should be viewed as the acquisition of a single, unitary, skill. The computational simulations reported below extend previous work developing the Chunk-Based Learner (McCauley & Christiansen, 2011, 2014) to show in detail how a unitary model of parsing and production may nonetheless exhibit behavioral parsing-production asymmetries. comprehension and production, this account appears to be challenged by an important empirical observation: the asymmetry often observed between comprehension and production in language acquisition. Although children in specific cases may exhibit adult-like production of sentence types that they do not appear to fully comprehend (cf. Grimm, Muller, Hamann, & Ruigendijk, 2011), their comprehension abilities generally appear to run ahead of their production skills (e.g., Fraser et al., 1963). But how can such asymmetries arise, if comprehension and production are as closely intertwined as we have proposed? In the next section, we consider a computational model compatible with the Now-or-Never bottleneck, and report new simulations that demonstrate how a com prehension–production asymmetry might arise within a system that unifies these two fundamental components of language use. A unified model of production and comprehension Inspired by the restrictions imposed by the Now-orNever bottleneck, McCauley and Christiansen (2011, 2014) implemented a computational model of language acquisition, the Chunk-Based Learner (CBL), which provides a unified approach to aspects of comprehension and production. The model aims to recreate the early language production of individual children given the linguistic input to which they have been exposed. Individual models are created for specific children in the CHILDES database (MacWhinney, 2000). The input to comprehension is the caregiver speech to the individual child, and production involves reproducing the utterances of that particular child. Thus, each CBL model simulates language acquisition in a single child, with child-directed speech (or speech spoken in the presence of the child) as input for comprehension and the child’s own utterances as the target for the model’s productions. The CBL model acquires item-based knowledge of the language in a purely incremental fashion. It learns online, using peaks and dips in ‘‘backwards” transitional probabilities between words to chunk words together as they are encountered, incrementally building up a ‘‘shallow parse” as each incoming utterance unfolds. By storing the word sequences that it groups together, the model gradually builds up an inventory of chunks consisting of one or more words – a ‘‘chunkatory” – which forms the basis for both comprehension and production. As it passes through the corpus, the CBL model attempts to reproduce each child utterance using only the chunks and distributional information it has acquired up to the point at which the child produced that particular utterance. Crucially, the very same chunks and distributional information used during production are also employed incrementally to build a ‘‘shallow parse” of utterances produced by caregivers. Thus, language acquisition in CBL 3 As we discuss below, the comprehension–production asymmetry may be somewhat overdetermined. There are additional, plausible, auxiliary assumptions that can also yield the asymmetry. Our contribution here is to point out that an asymmetry can arise directly from the operation of a unitary model, without depending on such auxiliary assumptions. N. Chater et al. / Journal of Memory and Language 89 (2016) 244–254 247 involves the simultaneous acquisition of two skills, concerning parsing and producing language: the model learns to become better at both skills, using the same chunkatory and the same distributional information within a unitary system. Moreover, the model’s emphasis on multiword chunks is consistent with evidence highlighting the importance of such units in both child and adult language comprehension and production (e.g., Arnon & Clark, 2011; Arnon & Snider, 2010; Bannard & Matthews, 2008; Janssen & Barber, 2012; Reali & Christiansen, 2007). CBL, in its present form, is limited to purely distributional learning and processing, involving combinations of words in individual utterances, and does not extract meaning from the utterance—hence, we will speak of CBL embodying a unified model of parsing and production, rather than comprehension and production. However, given the broader theoretical framework of the Now-orNever bottleneck (Christiansen & Chater, in press), we envisage that the kind of chunking the model performs (roughly akin to shallow parsing) will be supplemented by contextual top-down information (e.g., tied to semantic and pragmatic knowledge) to assist in full-blown language comprehension. This allows children to arrive at some rudimentary understanding of grammatical constructions they have not yet mastered in full (and therefore cannot use effectively in production). Through chunking, the model can arrive at an item-based shallow parse of an utterance; theoretically, a child might then use such a representation in conjunction with semantic and pragmatic information to arrive at a ‘‘good enough” interpretation of the utterance (Ferreira, Bailey, & Ferraro, 2002). Just as CBL does not model the extraction of meaning from linguistic input in comprehension, production in CBL does not address the process of deciding on the meaning to be conveyed. Instead the focus is on retrieving and sequencing words and chunks in the appropriate order to reconstruct the target child’s utterances. Thus, the model captures only some aspects of language comprehension and production. Yet this restricted model is sufficient to explore how behavioral asymmetries can arise through differing task demands, despite the use of the very same representations and cognitive mechanisms (i.e., the same chunks and distributional statistics) during both comprehension and production. Here, we expand on previous CBL simulations (McCauley & Christiansen, 2011, 2014) to explore how asymmetries between comprehension and production may arise, even in an entirely unified model of parsing and production. We first briefly outline the operation of 4 CBL was created to directly embody the twin constraints of incremental processing and on-line learning within a unified model of comprehension and production (McCauley & Christiansen, 2014). Other computational models have related properties. For example, Bod’s (2009) Unsupervised Data-Oriented Parsing provides an integrated, though non-incremental, model of production and comprehension in which fragments of current input are learned on-line, and used to parse or produce later sentences. Another example is the dual-mechanism model of Chang, Dell, and Bock (2006), who use a unified connectionist neural network architecture to simulate production and comprehension (in the form of next word prediction), but without being fully online and using a small artificial input corpus. Importantly, CBL has not previously been used to explore the comprehension/production asymmetry in a quantitative way, as in the present paper. the CBL model before reporting the outcome of a new set of simulations using analyses in which comprehension and production performance are scored according to the same metric. We then show, using natural performance measures, that the model exhibits better comprehension than production performance across sets of randomly selected test utterances, despite its unified architecture. Moreover, as CBL’s experience with processing language increases, the difference between comprehension and production performance decreases, as is also observed in child language acquisition.