com.kotlinnlp.neuralparser.parsers.transitionbased.models.arcstandard.td.BiRNNTDArcStandardParserModel.kt Maven / Gradle / Ivy

Go to download
/* Copyright 2017-present The KotlinNLP Authors. All Rights Reserved.
 *
 * This Source Code Form is subject to the terms of the Mozilla Public
 * License, v. 2.0. If a copy of the MPL was not distributed with this
 * file, You can obtain one at http://mozilla.org/MPL/2.0/.
 * ------------------------------------------------------------------*/

package com.kotlinnlp.neuralparser.parsers.transitionbased.models.arcstandard.td

import com.kotlinnlp.linguisticdescription.language.Language
import com.kotlinnlp.neuralparser.language.CorpusDictionary
import com.kotlinnlp.neuralparser.parsers.transitionbased.models.ScorerNetworkConfiguration
import com.kotlinnlp.neuralparser.parsers.transitionbased.templates.parsers.birnn.ActionsScorerNetworkBuilder
import com.kotlinnlp.neuralparser.parsers.transitionbased.templates.parsers.birnn.simple.BiRNNParserModel
import com.kotlinnlp.simplednn.core.functionalities.activations.ActivationFunction
import com.kotlinnlp.simplednn.core.layers.LayerType
import com.kotlinnlp.simplednn.core.neuralnetwork.NeuralNetwork
import com.kotlinnlp.syntaxdecoder.transitionsystem.state.scoreaccumulator.ScoreAccumulator

/**
 * The parser model for the ArcStandard parser based on the BiRNN with Transition+Deprel actions scorer.
 *
 * @property language the language within the parser works (default = unknown)
 * @property scoreAccumulatorFactory a factory of score accumulators
 * @property corpusDictionary a corpus dictionary
 * @property wordEmbeddingSize the size of each word embedding vector
 * @property posEmbeddingSize the size of each pos embedding vector
 * @param biRNNConnectionType the recurrent connection type of the BiRNN used to encode tokens
 * @param biRNNHiddenActivation the hidden activation function of the BiRNN used to encode tokens
 * @param biRNNLayers number of stacked BiRNNs
 * @param scorerNetworkConfig the configuration of the scorer network
 */
class BiRNNTDArcStandardParserModel(
  language: Language = Language.Unknown,
  scoreAccumulatorFactory: ScoreAccumulator.Factory,
  corpusDictionary: CorpusDictionary,
  wordEmbeddingSize: Int,
  posEmbeddingSize: Int,
  biRNNConnectionType: LayerType.Connection,
  biRNNHiddenActivation: ActivationFunction?,
  biRNNLayers: Int,
  scorerNetworkConfig: ScorerNetworkConfiguration
) : BiRNNParserModel(
  language = language,
  scoreAccumulatorFactory = scoreAccumulatorFactory,
  corpusDictionary = corpusDictionary,
  wordEmbeddingSize = wordEmbeddingSize,
  posEmbeddingSize = posEmbeddingSize,
  biRNNConnectionType = biRNNConnectionType,
  biRNNHiddenActivation = biRNNHiddenActivation,
  biRNNLayers = biRNNLayers) {

  companion object {

    /**
     * Private val used to serialize the class (needed from Serializable)
     */
    @Suppress("unused")
    private const val serialVersionUID: Long = 1L
  }

  /**
   * The neural network used by the ActionsScorer to score transitions.
   */
  val transitionScorerNetwork: NeuralNetwork = ActionsScorerNetworkBuilder(
    inputSize = 4 * this.deepBiRNN.outputSize, // the input is a tokens window of 4 elements
    inputType = LayerType.Input.Dense,
    outputSize = 4, // Shift, Root, ArcLeft, ArcRight
    scorerNetworkConfig = scorerNetworkConfig
  )

  /**
   * The neural network used by the ActionsScorer to score deprels.
   */
  val deprelScorerNetwork: NeuralNetwork = ActionsScorerNetworkBuilder(
    inputSize = 4 * this.deepBiRNN.outputSize, // the input is a tokens window of 4 elements
    inputType = LayerType.Input.Dense,
    outputSize = this.corpusDictionary.deprelTags.size + 1, // Deprels + Shift
    scorerNetworkConfig = scorerNetworkConfig
  )
}