The Now-or-Never bottleneck in language processing that Christiansen & Chater (C&C) propose has a great deal of evidence to support it. Like all cognitive processes, language processing must be implemented in neurons, and the bottleneck is a neural one. Signals from the environment must be processed by neurons, and those neurons must keep a memory trace of those signals or they will be lost. Moreover, any processing mechanism must not only be implemented by the behaviour of neurons, but in the case of language, the process must be learned by those neurons.
Neural memory comes in several forms. Neurons spike propagating signals across their synapses to post-synaptic neurons taking tens of milliseconds. Neurons can be wired into cell assemblies (Hebb Reference Hebb1949) that can persistently fire for seconds. Synaptic weights can be modified for seconds to minutes by means of short-term potentiation (STP), or for days, months, or longer, through long-term potentiation (LTP). The formation of a cell assembly, by potentiation, can form a circuit that can last indefinitely. When that long-term memory is activated by a cascade of neural firing in the cell assembly, the long-term memory is also an active short-term memory.
When a sentence is parsed, either in speech or in text, the parsing is generally done in one pass. This single pass can be seen in eye-tracking evidence, especially when repairs are needed (Just & Carpenter Reference Just and Carpenter1980). One pass parsing is typically simulated with a stack, but a memory-based mechanism (Lewis & Vasishth Reference Lewis and Vasishth2005) can eliminate the need for a stack. A memory-based parsing mechanism has been implemented in a neural parsing model (Huyck Reference Huyck2009), with the persistence of the cell assembly showing the strength and duration of the memory. I am unaware of any existing simulated neural mechanism for backtracking in parsing.
One important aspect of eliminating the stack in parsing is that it reduces the need for binding. Binding is another type of neural memory mechanism that, although needed in standard computational models, is typically overlooked. In a standard program, if a variable is assigned a value, the two are bound. Binding is usually a primitive operation so it is ignored. Binding in a neural system is more difficult because it is not primitive. There are various binding mechanisms; synchronous firing is most widely used in the literature (Fuster & Alexander Reference Fuster and Alexander1971). Two bound assemblies fire in roughly the same firing pattern, while another pair (or more) can be bound in a different pattern. Synchronous binding requires the neurons to continue firing. Moreover, only a small number of patterns can be supported simultaneously, so there are a limited number of bindings; the bound neurons do not all fire at the exact same time, so separate patterns must be quite distinct. Another option is to bind with STP. This method has neither of these limits, with a much larger number of bindings supported and the duration being up to minutes; it does, however, take longer to form. Binding can also be done with LTP, but this shades into permanent associative memory.
When people or computer systems process language, it is faster and safer to avoid binding. When binding is necessary, lower-level processing is likely to use synchrony. Higher-level processing is likely to use STP. So the speech signal uses synchrony; neurons representing the prime formants fire synchronously in the auditory cortex (Eggermont Reference Eggermont2001). The simulated neural parser (Huyck Reference Huyck2009) uses STP for binding the slots in the neural implementation of verb frames associated with sentences. These can be used immediately after sentence processing to retrieve the meaning of the sentence, but they are gradually erased by the STP fading. The neurons that support the binding are reused later for processing other sentences.
Finite-state automata (FSA) do not require binding. The evidence from text engineering to support the bottleneck is that the Message Understanding Competitions for Text Extraction (Appelt et al. Reference Appelt, Hobbs, Bear, Israel, Tyson and Ralescu1993) converged on an FSA cascade to solve the problem of processing text. One automaton separated words, a second categorised them lexically, a third did simple phrase parsing, and a fourth combined phrases. These could be run in a cascade, and perhaps a cascade is the basic mechanism that the brain uses.
As C&C note, the bottleneck also has ramifications for learning. First, the whole language cascade (whatever that may be) is being learned simultaneously. Initially, low-level phenomena, such as morphemes, are learned. Later, larger systems such as simple phrase grammars begin to be learned, but the lower-level systems are still being developed. We do not know how these biological neural systems work, much less how they are learned. One mechanism may be that things are being learned and cell assemblies (CAs) are formed; CAs can be connected to form FSA. Binding may be involved initially, and the synapse can then be modified to combine CAs into FSA; STP can support reverberation, which can then lead to LTP. Although one finite-state automaton in the cascade is being learned, both FSA above and below it can be learned so that the whole system continues to improve.
At the highest level, dialogue and above, the bottleneck begins to disappear. Rich cognitive maps support this kind of processing, and memory is formed mostly through LTP and CA circuit dynamics. Since these CAs can persistently fire, and the circuits can be reactivated using associative memory, it is possible to remember large amounts of things. (For example, I can still remember some of the dialogue from the movie I saw this weekend, and the plot.)
There is solid support for the Now-or-Never bottleneck in language processing, although the bottleneck's duration is reduced as the signal passes through stages of language processing. The distributed nature of neural processing supports multiple stages in processing, and the simultaneous learning of these stages. Processing and learning is implemented in neurons, although CA dynamics and binding issues are often not considered by researchers. By expanding understanding and modelling at the neural level, we can better understand language processing, and we can construct more robust language processing systems.
You have
Access
The Now-or-Never bottleneck in language processing that Christiansen & Chater (C&C) propose has a great deal of evidence to support it. Like all cognitive processes, language processing must be implemented in neurons, and the bottleneck is a neural one. Signals from the environment must be processed by neurons, and those neurons must keep a memory trace of those signals or they will be lost. Moreover, any processing mechanism must not only be implemented by the behaviour of neurons, but in the case of language, the process must be learned by those neurons.
Neural memory comes in several forms. Neurons spike propagating signals across their synapses to post-synaptic neurons taking tens of milliseconds. Neurons can be wired into cell assemblies (Hebb Reference Hebb1949) that can persistently fire for seconds. Synaptic weights can be modified for seconds to minutes by means of short-term potentiation (STP), or for days, months, or longer, through long-term potentiation (LTP). The formation of a cell assembly, by potentiation, can form a circuit that can last indefinitely. When that long-term memory is activated by a cascade of neural firing in the cell assembly, the long-term memory is also an active short-term memory.
When a sentence is parsed, either in speech or in text, the parsing is generally done in one pass. This single pass can be seen in eye-tracking evidence, especially when repairs are needed (Just & Carpenter Reference Just and Carpenter1980). One pass parsing is typically simulated with a stack, but a memory-based mechanism (Lewis & Vasishth Reference Lewis and Vasishth2005) can eliminate the need for a stack. A memory-based parsing mechanism has been implemented in a neural parsing model (Huyck Reference Huyck2009), with the persistence of the cell assembly showing the strength and duration of the memory. I am unaware of any existing simulated neural mechanism for backtracking in parsing.
One important aspect of eliminating the stack in parsing is that it reduces the need for binding. Binding is another type of neural memory mechanism that, although needed in standard computational models, is typically overlooked. In a standard program, if a variable is assigned a value, the two are bound. Binding is usually a primitive operation so it is ignored. Binding in a neural system is more difficult because it is not primitive. There are various binding mechanisms; synchronous firing is most widely used in the literature (Fuster & Alexander Reference Fuster and Alexander1971). Two bound assemblies fire in roughly the same firing pattern, while another pair (or more) can be bound in a different pattern. Synchronous binding requires the neurons to continue firing. Moreover, only a small number of patterns can be supported simultaneously, so there are a limited number of bindings; the bound neurons do not all fire at the exact same time, so separate patterns must be quite distinct. Another option is to bind with STP. This method has neither of these limits, with a much larger number of bindings supported and the duration being up to minutes; it does, however, take longer to form. Binding can also be done with LTP, but this shades into permanent associative memory.
When people or computer systems process language, it is faster and safer to avoid binding. When binding is necessary, lower-level processing is likely to use synchrony. Higher-level processing is likely to use STP. So the speech signal uses synchrony; neurons representing the prime formants fire synchronously in the auditory cortex (Eggermont Reference Eggermont2001). The simulated neural parser (Huyck Reference Huyck2009) uses STP for binding the slots in the neural implementation of verb frames associated with sentences. These can be used immediately after sentence processing to retrieve the meaning of the sentence, but they are gradually erased by the STP fading. The neurons that support the binding are reused later for processing other sentences.
Finite-state automata (FSA) do not require binding. The evidence from text engineering to support the bottleneck is that the Message Understanding Competitions for Text Extraction (Appelt et al. Reference Appelt, Hobbs, Bear, Israel, Tyson and Ralescu1993) converged on an FSA cascade to solve the problem of processing text. One automaton separated words, a second categorised them lexically, a third did simple phrase parsing, and a fourth combined phrases. These could be run in a cascade, and perhaps a cascade is the basic mechanism that the brain uses.
As C&C note, the bottleneck also has ramifications for learning. First, the whole language cascade (whatever that may be) is being learned simultaneously. Initially, low-level phenomena, such as morphemes, are learned. Later, larger systems such as simple phrase grammars begin to be learned, but the lower-level systems are still being developed. We do not know how these biological neural systems work, much less how they are learned. One mechanism may be that things are being learned and cell assemblies (CAs) are formed; CAs can be connected to form FSA. Binding may be involved initially, and the synapse can then be modified to combine CAs into FSA; STP can support reverberation, which can then lead to LTP. Although one finite-state automaton in the cascade is being learned, both FSA above and below it can be learned so that the whole system continues to improve.
At the highest level, dialogue and above, the bottleneck begins to disappear. Rich cognitive maps support this kind of processing, and memory is formed mostly through LTP and CA circuit dynamics. Since these CAs can persistently fire, and the circuits can be reactivated using associative memory, it is possible to remember large amounts of things. (For example, I can still remember some of the dialogue from the movie I saw this weekend, and the plot.)
There is solid support for the Now-or-Never bottleneck in language processing, although the bottleneck's duration is reduced as the signal passes through stages of language processing. The distributed nature of neural processing supports multiple stages in processing, and the simultaneous learning of these stages. Processing and learning is implemented in neurons, although CA dynamics and binding issues are often not considered by researchers. By expanding understanding and modelling at the neural level, we can better understand language processing, and we can construct more robust language processing systems.
ACKNOWLEDGMENT
This work was supported by the Human Brain Project Grant 604102, Neuromorphic Embodied Agents that Learn.