OpenDRC: An Open-Source Implementation
of the DRC Model

OpenDRC is a open-source implementation of the Coltheart, Rastle, Perry, Langdon and Ziegler (2001) model of text-to-speech conversion in the human brain. The model is designed to explain the computational process of skilled readers in basic reading tasks. Further detail of the model can be found in the Coltheart, et. al. (2001) Psychological Review paper describing the DRC model and Steven Saunder's paper describing the difference between DRC1.0 and DRC1.2 (see link above).

The acronym, DRC (Dual Route Cascaded), signifies the two main concepts of the model: it is a dual route and cascaded model. The full model concept incorporates three routes: Lexical, SubLexical (or Grapheme-to-Phoneme Conversion (GPC)) Route and the Semantic Route (not implemented in either implementation). Each of the implemented routes consists of a number of layers and each layer a number of individual nodes. The nodes are the smallest structures in the model and represent such things as letter features in the Feature-Level(FL), letters in the Letter Level(LL), words in the Orthographic Input Lexicon(OIL), phonemic words in the Phonographic Output Lexicon(POL) and phonemes in the Phoneme Buffer(PB).

The DRC model processes input text strings, whether words known to the model or non-words, using both a lexical and a sublexical pathway. The lexical pathway is the main route for processing of known words and transverses the Orthographic and Phonological lexicons. The sublexical pathway is the route for processing unknown words or non-words and transverses the Grapheme-to-Phoneme Conversion layer. The model has cascaded activation through the layers where one layer's activations are fed through to the next (and previous) layer(s) in subsequent processing cycles and thus cascades through the model. The model layers interact through excitatory or inhibitory activations and nodes within a layer may laterally inhibit each other.

The OpenDRC implementation, like DRC1.2, has many variables in addition to the default language parameters for controlling the functioning of the model. These parameters control the degree to which one layer's activation can excite or inhibit an adjoining layer as well as many other characteristics of the model that affect the processing of an input text string. The DRC model can also be lesioned to model the effects of human brain injuries as well as to inspect the effects of various input classes (pseudohomophones, word-length, neighbourhood density and so on) on the processing speed of the model.

The OpenDRC project was started to proivde the research community a open-source version of the DRC1.2 model which could be modified by individual researchers. OpenDRC also allows researchers to inspect the workings of the computational model. This allows all researchers to modify the model in their own direction, and check for and debug suspected implementation errors of the model with which they are testing. During the construction of OpenDRC one such error was discovered: The Unsupported Decay Bug in the Windows and Linux versions of DRC1.2.

Verification of OpenDRC

OpenDRC has been tested against the DRC model implementations (versions DRC1.2 and DRC1.2.1) for simple (no masking or inter-word interference) batch processing of input words. The tests all show near identical outputs with the exception of a handful of vocabulary words and non-words.

The testing process involves the use of three batch sets: The entire DRC1.2/english1.1.6 and DRC1.2.1/english1.1.7 vocabulary (7978 words known to the model), 1000 non-words from the ARC Non-Word Database, and a set of 80 pseudohomophones from McCann&Besner (1987). In these three batch tests the DRC1.2(.1) and OpenDRC output RT, STATS, Parameters and ACTS files are compared (using the OpenDRC utility DRCTest) to determine compliance between OpenDRC and DRC1.2(.1).

The Run-Time(RT) file comparisons indicate a correlation of 0.998 or better on each test of the vocabulary, pseudohomophones and the non-word lists. OpenDRC and DRC1.2(.1) differ in RunTimes for one vocabulary word {POOH(100/101 cycles)} and interpret the correctness-of-translation differently for three over-long (DRC was defined to maximum word length of 8 characters) vocabulary words {SCROUNGED, STRAIGHTS, and STRENGTHS}, two ARC NonWords {SCROOLED,THROUGED} and no Pseudohomophones. DRC1.2 treats translations of over long words (>8 characters) WRONG and OpenDRC flags them as CORRECT because, although the translation is WRONG for the exterally visible word, it is CORRECT for the internally visible word. DRC1.2 tests the full word against the corresponding pronuciation, OpenDRC compares the truncated word with the vocabulary.

	English 1.1.6			English 1.1.7
	≥1% RT	≥1.5x10-6 Actsb	Translationc	≥1% RT	≥1.5x10-6 Actsb	Translationc
English 1.1.6 Vocabulary (7978 Words)	1/3a	7	13	1/3a	7	3
ARC NonWord DB (1000)	0/2a	8	0	0/2a	7	0
Pseudo-Homophones (80)	0	0	0	0	0	0

Table 1: OpenDRC test results on language database English1.1.7(DRC1.2.1) and English1.1.6(DRC1.2).

Notes:

1. RT errors are expressed as A/B where A is the number of inputs with RT differences and B is the number of inputs with CORRECT/WRONG differences, or in the case of ARC Nonwords B is the number of words that also hit the LOWAC translation termination condition. In this case DRC1.2(.1) considers the word's translation CORRECT whereas OpenDRC classifies them as WRONG.
2. All activation errors are ≤ 4.0x10^-6 and may appear anywhere within the ACTS files.
3. A word/nonword whose translation differs in its ACTS file from DRC1.2(.1) to OpenDRC.

The STATS file comparison of the OpenDRC results closely agrees with those of DRC1.2.1 with the difference attributable to the above RT differences.

The Parameters file comparison showed only formatting differences and additional comments in OpenDRC's output.

The ACTS files (detailed report of node activations) were compared both for structural similarity and activation values. The activation value test is a much more stringent test that the RT cycle-time test. The Error-Threshold for individual activations is set at 0.00000150 and 0.00000375 for layer totals. Of the 7978 vocabulary words only ten words failed this test {MYRRH, POOH, YEAH} due to a translation difference between DRC1.2.1, which did not find a translation, and OpenDRC which did, and over-long vocabulary words {SCRATCHED, SCREECHED, SCRUNCHED, SPLOTCHED, SQUELCHED, STAUNCHED, STRETCHED}. Of the 1000 ARC Non-Words only eight failed the test {ANK, CHUNCHED, CRUTCHED, SCROOLED, STRUMPED, THEZ, THROUGED, TRUNCHED} which failed because activation differences exceeded the thresholds but were less than 0.000002(individual) and 0.000006(totals). No Pseudohomophones failed the ACTS tests.

Thus OpenDRC and DRC1.2(.1) differ in RT values in only one of the 9058 test cases. They classify five(5) words differently: CORRECT/WRONG. OpenDRC differs from DRC in 14 or 15 test cases for activations (which are all ≤4.0x10^-6 and about 100,000 times smaller than typical activations). Finally, OpenDRC differs from DRC in ≤13 translation differences out of the 9058 test cases.

Download OpenDRC

OpenDRC Suite

The OpenDRC project contains a number of programs to help the researcher work with the OpenDRC (and DRC1.2.1) models.

OpenDRC

OpenDRC is the main program. It can be run at the command line in a similar fashion to DRC1.2/DRC1.2.1. The OpenDRC command line parameter list has been implemented to be a superset of that of DRC1.2. There have been additions to the list including "--neighbourhood" to display the orthographic neibourhood of the given words; which is dependent of the model's vocabulary. Other parameters have been added as well as the model's operating parameter list. For example, inter-layer thresholding has been added to OpenDRC1.02.

DRCTest

This program is a helper program used to test for difference in the <batchname>drc output directories produced by either OpenDRC or DRC1.2/DRC1.2.1. This program, while compairing any two output directories, will find any differences in the corresponding output RT, STATS, Parameter, or ACTS files produced. The RT, STATS and Parameter files will be tested in a similar fashion to a standard Linux diff command. Each individual ACTS file will be tested against it's corresponding ACTS file in the other directory. However, the test will look for structural differences in these files (added/changed or deleted lines for example) but will ignore any activation differences that do not exceed given thresholds. These thresholds are for the individual activation values (default:1.5x10^-6) or activation total values (default:3.75x10^-6). This is done because OpenDRC and DRC1.2 calculations appear to suffer from floating point roundoff differences. Since these differences are minimal they can safely be ignored. However, command line parameters are available for the researcher to adjust these two thresholds.

DRCFilter

DRCFilter allows the researcher to extract the activation information from one or more ACTS files and will output this information in a form suitable for inclusion in MS Excel/LibreOffice-Calc or Matlab/Octave so that the activation history can be graphed. Separate Matlab routines are also included in OpenDRC1.02 that will display the activations in a graph for the Matlab user. These can be used as a starting point for researchers own Matlab programs.

SourceForge.net

The OpenDRC project is hosted on the SourceForge.net website. The SourceForge summary page for OpenDRC contains specific information on the project.

The author can be reached at