CLEAR

Brief: Computer LEarning ARchitecture
Jump To: Parent Description
Code: GitHub

  • CLEAR is a great way to have books, manuals, websites, etc, read to you, allowing you to pause, quit, resume, and filter out nonsense. Clear uses the Festival text-to-speech system and ccp to do this. It is very useful for studying. For instance, while browsing a researchers website in w3m-el, you can select a region over all of their publications and execute M-x clear-queue-all-links or "\C-c\C-mc" to queue all of their papers. A message containing a command to queue the links is then sent from Emacs-Unilang-Agent to UniLang, which sends it to CLEAR. The queued links are stored to the current readlist. Whenever you have a chance then, you tell CLEAR to resume reading and it reads you books. It sorts the reading list by topical dependency.

    The objective is to model the user's reading flux at the sentence level. Conceptual understand may then be modelled through analysis of the sentences. A lot of data on the entire process is recorded. The purpose is so that the computer has another modality, in addition to things like expected background knowledge, apparent knowledge, to model which axioms the user is familiar with. (Eventually the Textbook Knowledge Formation tool chain we are assembling with the Chess Analysis Knowledge Base project will be used to model fields axiomatically as opposed to the document level granularity currently used by CLEAR.) By modelling the reading flux, a much better mental model can be created and the system can behave contingently. This information is critical to many tasks, such as:

    • Assessing the user's background knowledge.
    • Verifying familiarity with some tradecraft.

    The motivation for CLEAR is inspired in part by the following information from the JAVELIN project: http://www.lti.cs.cmu.edu/Research/JAVELIN/

    Utility-based Information Fusion. Any item of information I can be assigned a value representing its utility to analyst A with respect to task context T and question Q. The utility value can be used to rank the possible answers in a manner inspired by Maximal Marginal Relevance:

    U = Argmaxk[F(Rel(I,Q,T),Nov(I,T,A),Ver(S,Sup(I,S)),Div(S),Cmp(I,A)),Cst(I,A)]

    Essentially all information items (facts, links, inferred relations, etc.) in consideration for fusion into an answer may be ranked in utility to the analyst as a function of:

    • Rel: relevance to the requested information
    • Nov: novelty (likelihood that the analyst does not already know it)
    • Ver: veracity (of source S) and support for conclusion I within S
    • Div: source diversity (analyst may want contrasting or reinforcing views)
    • Cmp: comprehensibility of information by the analyst (one can assume a uniform distribution until the system learns otherwise from analyst feedback)
    • Cst: expected cost (e.g. time) for the analyst to assimilate the information.

    The way CLEAR does this is by monitoring the users readers, and deriving belief models of the user's state (such as awareness, eye-position, whether they were listening (in the case of auditory media) etc), as these reader applications were running. Currently only reading is implemented, as I don't have a suitable eye-tracker.

    In addition to reading, currently CLEAR can quiz the user, as in this example:

    /var/lib/myfrdcsa/codebases/internal/clear $ clear -q lists/camo.rl 
    Readlist is lists/camo.rl
    Initializing TTS engine...
    festival: no process killed
    server    Tue Dec 28 20:10:06 2004 : Festival server started on port 1314
    Reviewing library/us-army-field-manuals/extracted/20-3/US ARMY FM 20-3 Camoflage Concealment And Decoys/ch4.PDF...
    client(1) Tue Dec 28 20:10:08 2004 : accepted from pc-sonicparts
    
    Question 0
    Make optimum use of concealed routes, hollows, gullies, and other terrain
    features that are dead-space areas to enemy observation and _____
    positions.
    % fighting
    Incorrect! (firing)
    
    Question 1
    Although an enemy's use of radar and _____ aerial recon hinders operations
    at night, darkness remains a significant concealment tool.
    % visual
    Incorrect! (ir)
    
    Question 2
    When natural _____ and concealment are unavailable or impractical, the
    coordinated employment of smoke, suppressive fires, speed, and natural
    limited-visibility conditions minimize exposure and avoid enemy fire sacks.
    % cover
    Correct! (cover)
    
    Question 3
    Designate concealed _____ for movement into and out of an area.
    % routes
    Correct! (routes)
    
    Question 4
    A trade-off, however, usually exists in _____ of a slower rate of movement
    when using these types of routes. 4-18.
    % this
    Incorrect! (terms)
          
    One can see how Question 4 is not as useful as the others, but overall, I find the system to be very effective in liu of a real system. If the shoe fits...

    This system used to be called: Correctly Learning An Individual's Reasoning Visually Or Yielded Assuming Normal Cognitive Execution (CLAIRVOYANCE), but this is just sort of inaccurate. CLEAR is much, more, well, clear ;)

    CLEAR is primarily useful for reading documentation, papers, books, websites, digital library content, and other sources that are of practical use to the project, so that we can get much more done, especially while resting or moving. A critical feature is the ability to read documentation for installed packages, because mastering newly installed software is critical to the FRDCSA mission goal. It also does very well with important documents, like the above survival and preparedness information (as the recent tsunami confirms). Lastly, and most important, it allows us to implement a system of instruction so that we can help to train project members in various capacities and verify their progress, without being burdensome at all - but rather entertaining. This can be combined with CRITIC for collaborative filtering based ratings to enhance the value of the information. CLEAR in conjunction with DigiLib is also used by job-search to ensure familiarity with position requirements. CLEAR uses functionality from CoAuthor to compose custom reading specific to users' tested proficiency and known reading history.