Transcript Foresight - Unique Media TV

Cognitive Systems
Foresight
Language and Speech
Cognitive Systems
Foresight
Language and Speech
• How does the human system organise itself, as a neuro-biological
system, to integrate top-down and bottom-up information during
language interpretation and production? How does this relate to the
organisation of current natural language processing systems?
• Assumption that there is an organisation in this way: perhaps only
bottom-up in the human system.
• Some of the system is pre-wired and some is learnt during early
plasticity, which affects how the system organises itself.
• What sort of information does the engineer need to know about the
structure? Neuro-imaging uses? Blood flow is only one level of
information: what about the neural mechanisms?
Cognitive Systems
Foresight
Language and Speech
• Evidence of sub-processes running in parallel from fMRI, rather than the
single stream approach.
• Importance of feedback processes as well as feedforward processes:
not seen in current engineering processes. Birdsong depends critically
on feedback.
• Don’t yet have the contextual understanding of communication situation,
which is a major limiting factor on speech recognition.
• Where do human systems asymptote: not 100 %.
Cognitive Systems
Foresight
Language and Speech
• How far are the characteristics of human processors determined by the
statistical properties of the speech input and how do these relate to
current statistical techniques used in automatic speech recognition?
• There must have been some co-evolution of the production and
perception systems.
• Are current techniques more sophisticated (HMMs) in statistical
processes than the brain to attempt to make up for a lack of higher
understanding? Some disagreement over this.
• Brain is trained on moving segments, rather than static segments, so
perhaps a pattern recognition approach is fine, but the training is the
problem.
Cognitive Systems
Foresight
Language and Speech
• What problems must we solve to develop human-computer interfaces which
demonstrate human levels of robustness and flexibility?
• Theory of mind. Some engineering systems are ‘single level’ without any
understanding of higher levels.
• No discussion of the grounding of language. Use of analogy: can a computer
understand language without a grounding of language?
• Solving the problem of speech is not the same as solving the problem of
language.
• Communication between brains and communication within brains: what can one
teach about the other?
• Why do only humans have language? Ability of humans to replicate samples in
correct temporal scales on short timescales. Representation of time. Some
dissent on this point.
• Important to get context right, so other modalities (e.g. body language) are
important. These can fill out areas where the speech recognition fails. Need for
sensor fusion.
Cognitive Systems
Foresight
Language and Speech
• Will the study of language in its interactive context lead to new
approaches to basic language processing, and to the design of dialogue
systems?
• Almost certainly: no grammars of language interaction and use, so we
need these, which will hopefully come from our study of natural
communication.
• Real language versus ideal language.
• Speech development, looking at young children. Small vocabularies but
good repair structures.