DOC PREVIEW
Improving the Reading Rate of Double-R-Language

This preview shows page 1-2 out of 6 pages.

Save
View full document
View full document
Premium Document
Do you want full access? Go Premium and unlock all 6 pages.
Access to all documents
Download any document
Ad free experience
View full document
Premium Document
Do you want full access? Go Premium and unlock all 6 pages.
Access to all documents
Download any document
Ad free experience
Premium Document
Do you want full access? Go Premium and unlock all 6 pages.
Access to all documents
Download any document
Ad free experience

Unformatted text preview:

Improving the Reading Rate of Double-R-Language Mary Freiman1 and Jerry Ball2 L3 Communications1 Air Force Research Laboratory2 [email protected], [email protected] Abstract This paper describes changes to a model of reading comprehension to improve its reading rate and bring it into closer alignment with human reading rates. The broader context of the research is development of language capable synthetic teammates that can be integrated into team training simulations. To use synthetic teammates in team training without detriment, we believe the synthetic teammates must be both functional and cognitively plausible. By functional, we mean that the synthetic teammate operates in real time, performs the task, and handles the range of linguistic inputs that are encountered. By cognitively plausible, we mean that the synthetic teammate adheres to well established cognitive constraints on human language processing—including the incremental and interactive processing of language at human reading rates. Achieving human reading rates in a cognitively plausible and functional model of reading comprehension is a research challenge that has not been met to date. Keywords: human language processing, reading rate, synthetic teammate, functional, cognitively plausible Introduction We are developing a model of reading comprehension called Double-R-Language (Ball, 2007; Ball, Heiberg & Silber, 2007). Double-R stands for Referential and Relational—two key dimensions of meaning that get grammatically encoded in English. The initial application of the reading model is development of a synthetic pilot for use in a three-person UAV simulation. The synthetic pilot flies the simulated UAV from a ground control station and will eventually communicate with a human navigator and photographer in the completion of reconnaissance missions. A prototype system has been developed (Ball, et al., 2009) using the ACT-R Cognitive Architecture (Anderson, 2007). The synthetic pilot prototype communicates with lightweight agent versions of the navigator and photographer developed outside ACT-R. The prototype communicates with the navigator and photographer using text chat and must be capable of reading and comprehending the messages it receives from them. The reading comprehension model is capable of incrementally processing linguistic inputs and generating linguistic representations of referential and relational meaning. These linguistic representations are interactively mapped into a non-linguistic representation of the objects and situations referred to in the linguistic input. The non-linguistic representation—called the situation model (cf. Zwann & Radvansky, 1998)—drives the task behavior of the synthetic pilot and determines when to communicate with the other teammates to acquire needed information. A significant challenge for the reading comprehension abilities of the model is input variability. A corpus of text chat communications that was collected in an experiment involving human subjects and the UAV simulation is full of variability in the form of linguistic input (see Table 1). For competent readers, misspelled words activate the intended lexical items because they contain many of the same letters and trigrams (Perea & Lupker, 2003). Hence, key requirements of the reading model include the ability to handle misspellings in input; the ability to separate perceptually conjoined units (e.g. separating punctuation from words as in ―He went.‖, but not ―etc.‖; separating words lacking spaces as in ―yougo‖ for ―you go‖); and the ability to recognize multi-word expressions (e.g. ―speed up‖) and multi-unit words (e.g. ―a priori‖, ―h-area‖). Table 1. Messages seen during a UAV simulation To satisfy these requirements, the model includes a word recognition subcomponent that uses ACT-R’s spreading activation mechanism to influence lexical item retrieval. The subcomponent maps orthographic input directly into DM representations without recourse to phonetic processing, although a phonetic mapping is not precluded. The model uses the spreading activation mechanism of ACT-R to retrieve words from the lexicon that are not an exact match to the input. Letters and trigrams in the input spread activation to the words containing those letters and trigrams in the mental lexicon. These processes and encodings are based on the Interactive Activation model of word recognition (McClelland & Rumelhart, 1981), with the addition of trigrams based on ―letter triples‖ (Seidenberg & McClelland, 1989). The subcomponent is embedded in the reading comprehension model as a whole; the effects of context and previous activation levels are taken into consideration when encoding each individual word (Freiman & Ball, 2008). The reading model also includes a verification stage to check the retrieved lexical item against the perceptual input. The verification stage aligns with the Activation-Verification model of Paap et al. (1982). It splits concatenated words in the input (e.g. ―yougo‖) to match the MESSAGE: VARIANT: i need to be beloe 3000 for f area i; beloe; f area effective radiu any requirements for altitde/speed? can yougo faster yet or is it stll 200 radiu altitde yougo; stll 67retrieved word (e.g. ―you‖), leaving a residual (e.g. ―go‖) for subsequent processing. If the retrieved lexical item is not a sufficiently close match to the input, the model treats the input as an unknown word. Even without considering the mapping of the linguistic representations into the situation model, the previous version of the reading model was much slower than humans in both cognitive processing time and real time performance. Adult readers read at a rate of 200-300 words per minute (Taylor, 1965; Carver 1973a; Carver 1973b). The average reading rate of the model—prior to the introduction of the changes described in this paper—was 96 words per minute (cognitive processing time), making it impossible to match the model’s performance against human performance. Since we are interested in building a model of reading comprehension that is cognitively


Improving the Reading Rate of Double-R-Language

Download Improving the Reading Rate of Double-R-Language
Our administrator received your request to download this document. We will send you the file to your email shortly.
Loading Unlocking...
Login

Join to view Improving the Reading Rate of Double-R-Language and access 3M+ class-specific study document.

or
We will never post anything without your permission.
Don't have an account?
Sign Up

Join to view Improving the Reading Rate of Double-R-Language 2 2 and access 3M+ class-specific study document.

or

By creating an account you agree to our Privacy Policy and Terms Of Use

Already a member?