The role of competition and timeframes: A commentary on Ambridge (2020)

Ambridge (2020) argues that language learning models do not need to postulate slot-and-frame patterns, but can rely solely on exemplar storage. This powerful idea does a good job of capturing the idea that language learning is heavily data-based. My own approach to language learning has a very similar emphasis. However, unlike Ambridge, my account has viewed patterns of overgeneralization as arising from a competition between rote and combination. Rote retrieval is required, even by adults, for forms that resist regular patterns. Combinatorial patterns arise through the abstraction or superimposition (MacWhinney, 1975b) process from this rich database, using algorithms much like the TiMBL. These patterns arise to control word order (MacWhinney, 1982), morphology (MacWhinney, 1978), meaning (MacWhinney, 1989), and phonology (Bybee & Beckner, 2015), when items are combined into phrases and utterances. Combinatorial patterns exist on three levels of abstraction (MacWhinney, 1975a, 2014). The lowest level is the item-based pattern (MacWhinney, 1975a). From these, feature-based patterns can be extracted (MacWhinney, 1982), and in some cases, general patterns (Perfors et al., 2010). Unlike Ambridge, however, I have assumed that these patterns are stored, perhaps through local topological organization. Back in 1978, our understanding of the brain and models gave us few cues about how this storage might happen. However, it now appears that this can be done through the development of self-organizing feature maps that take advantage of connectivity patterns in cortical layers (Li et al., 2007).

The crucial assumption in the Competition Model is that on-line processing involves a dynamic competition between rote and combination. Because of this, we (Bates & MacWhinney, 1982) have referred to the account as the Competition Model. Once one allows for competition between rote and combination, the objections raised by Ambridge against the storage of item-based patterns disappear. Consider an overgeneralization such as I poured the tub with water (Bowerman, 1988). In this case, item-based patterns related to verbs, such as fill, load, and cover give rise to a feature-based verb-group pattern of agent + action + goal + with + transferred. This pattern is overgeneralized to pour because of semantic similarity to the group and to fill in particular. Correction of this error arises from two sources. One is through strengthening of the agent + pour + transferred + goal item-based pattern for pour. The other is through the linking of pour to the feature-based verb-group pattern of agent + action + transferred + goal for verbs, such as dump, shovel, and stuff. Ambridge would explain this overgeneralization and subsequent recovery in terms of analogical processing. He invokes Occam’s Razor to argue against positions, such as the Competition Model, that postulate multiple processes. He claims that no paper ‘has made the case for a phenomenon that cannot be captured by a pure exemplar model, whether in the domain of language, or of learning and categorization more generally.’

In fact, there are several pervasive phenomena that are not captured by Ambridge’s model. The first phenomenon is the U-shaped learning of morphophonotactic patterns. Children acquire high frequency English irregular past tense verbs, such as went or fell, early on before learning the general combinatorial pattern of suffixing through combination with -ed. Once the combinatorial pattern arises, there is a period of competition between rote and combination that leads to variation. Eventually, combination dominates unless the rote form is strengthened by frequent exposure. This effect can even be seen for high frequency regulars such as wanted in terms of their greater speed of access and resistance to error as compared to low frequency regulars (Stemberger & MacWhinney, 1986). It is not clear how the analogy-only model can handle U-shaped patterns and differential processing of regulars without building in something that mimics the two processes in time (Kawamoto, 1994).

As MacWhinney (1978) noted, the competition between rote and combination is also revealed in on-line processing errors. On the morphological level, children and adults will correct overregularizations on-line by saying runned, uh ran. In such cases, combination first wins out in the horserace, but then a weaker rote form is eventually retrieved. Reformulations reflecting the horserace occur on other levels too. In descriptions of simple transitive pictures (MacWhinney & Bates, 1978), children (and some adults) might begin, on the level of utterance formulation, with the cat is getting the flower from the bunny in which the recipient is chosen as the perspective, but then they shift in midstream to taking the preferred perspective of the giver to formulate the bunny is giving the flower to the cat. The result of this competition between perspectives and verbs could be an utterance with retracing, such as the cat is get- uh the bunny is giving the cat a flower. On the lexical level, a child may at first call an animal a kitty and then correct to tiger. On the phonological level, a child may first produce banana as /nana/ with initial syllable deletion, but then correct to the full form. In each of these cases, a weaker or slower rote form competes with the quicker general pattern.

A third line of evidence involves our ability to adjust on-line to unfamiliar accents or languages. When first hearing a few sentences produced with a Scottish accent, I can fail to pick up the message, but then suddenly my frameworks for Scottish phonology become activated and things start to become clear. The same can occur when hearing, at some distance, some language, other than English, that I might know. At first I just hear sounds, but then the set of lexical and phonological processors for that language become activated and I can easily follow.

These three lines of evidence for a two-process or even three-process (rote, combination, and analogy) account share a common emphasis on the role of the multiple timeframes in processing and development (MacWhinney, 2005, 2015). The analogy-on-the-fly account fails to consider the dimension of time. Given enough time, we can retrieve weak memories, but language operates against a now-or-never bottleneck (Christiansen & Chater, 2016) that requires optimization (Anderson, 1990) of cortical organization to facilitate processing in real time. We do not need to imagine that this requires modular separation of rote from combination. Instead, combinatorial groups could be stored in self-organizing feature maps (Li et al., 2007). Patterns of this type in temporal cortex could be activated by combinatorial operations involving frontal cortex, based on the reciprocal connectivity between these areas.

Despite the omission of a consideration of competition and timescales, Ambridge’s paper makes an important contribution by focusing attention to the ways in which exposure to a rich input database can produce persistent exemplar storage and the extraction of patterns from this database.