The SARDSRN Sequential Parser

The networks have been trained on a corpus of 263 sentences with one embedded relative clause. Validation and testing are performed on the remaining 173 sentences from the grammar

The task is shift-reduce parsing. The networks read a sequence of input word representations into output patterns representing the parse results, which are syntactic constituents. At each time step, a copy of the hidden layer (or previous outputs for NARX networks) is saved and used as input during the next step, together with the next word. In this way each new word is interpreted in the context of the entire sequence so far, and the parse result is gradually formed at the output. Reductions are RAAM representations, which have been trained beforehand. The RAAM representations can be examined in the separate popup window by clicking the left mouse button on the output layer, which will send that representation through the decoder and display the RAAM components of the representation in the two assemblies labelled "Left" and "Right". Clicking the left mouse button on either of the decoder output assemblies (the ones above the "Left" and "Right" targets will further decode the representations. The initial RAAM representation is always saved, so it can be restored by clicking on the middle mouse button if another path through the decoding process is desired. Given this background, we can understand the components of the demo:

1. Main Window:
   1. Menubar:
      • RUN through all the examples; (The display rate is set at 1, so this will still show every step of the parsing process.)
      • STEP through a particular example; (Also the right mouse button)
      • CLEAR everything to start over again;
      • RESET the network (NOTE: this will randomize the weights; you will need to run the demo again to get the previous weights);
      • QUIT the demo.
      • An entry widget giving the current path for the grammar being investigated ("experiments/sardsrn/parse01_demo", by default). Using the cursor, you can change the file to "parse01_train" to see how the network performs on the training data.
   2. The Network: from top to bottom, left to right, the layers are:
      • input: the input words.
      • map: the SARDNET sequential SOM.
      • previous outputs: a layer to hold up to six prior outputs for NARX and SARDNARX.
      • previous hidden: holds the previous hidden layer for SRN and SARDSRN.
      • hidden: the hidden layer for all networks
      • output: the output where shifts and reductions are determined.
      • target: the proper targets.
      There are four networks that can be examined:
      • SRN (SRN on, SARD and NARX off): only the previous hidden layer is used.
      • SARDSRN (SRN on, SARD on, and NARX off): both the previous hidden layer and the map are used.
      • NARX (SRN on, SARD off and NARX on): only the previous six outputs are used.
      • SARDNARX (SRN, SARD and NARX all on): the map and the previous six outputs are used.
      The three components (SRN, SARD, and NARX) can be activated, tested, and trained separately or together via the checkbuttons below the menubar. For example, the network's performance without the SARD component can be tested by clicking on the checkbutton itself, thereby deactivating it. Clicking it (so that the checkbutton is activated) will reactivate the SARDNET input. For the purposes of the demo, a checkbutton labelled "Fulltest" in the RAAM window allows full testing of the network to be performed (see next item). If turned off, the network will run faster.
   3. A Status line showing the error signals from the SRN and SARD components of the network, and the combined error of the two components. This latter error also shows the average error per output unit per epoch upon completion of each epoch (normally just one epoch is needed for testing).
2. RAAM Popup Window: This window only shows the decoder part of the original RAAM network used to derive the partial parse representations. It is used here to allow recursive decoding of these representations, as well as displaying the statistics from "FullTest".

   This window also displays the statistics showing the average error per output unit upon decoding, as well as "Leaferror" and "Mismatch" statistics give additional, more stringent, measures of the network's performance, and are part of the test suite performed by the "FullTest" routine. "Leaferror" shows the leaf error, i.e., the error measured by descending through the leaves of the RAAM representations in addition to the output-target error. The "Mismatch" measure shows the number of actual lexical mismatches that occur as the leaves of the RAAM representations are unwound. If the RAAM representation at the output of the SARDSRN network only superficially encodes the parse result, then descending the RAAM representation tree will reveal that the noise in the leaves themselves is too high. See SARDSRN: A Neural Network Shift-Reduce Parser for details.

3. Inspect Map (SARD): This small window will display the vectors associated with each unit in the feature map (the big black square in the main window). Simply using the left mouse button over the feature map will display a vector representation, and the unit number and coordinates of the unit within the map. This is useful to get an idea of how the map is (self-)organized following training.

1. 0.00-0.25: black to dark to light blue
2. 0.25-0.50: light blue to olive
3. 0.50-0.75: olive to dark violet
4. 0.75-1.00: dark violet to white

Back to UTCS Neural Networks home page