Reads a CoNLL-2008 formatted file (containing semantic roles) and converts it to our own Metal format User: mihais Date: 5/5/15 Last Modified: 08/05/2020: Added the latest Metal format Update for Scala 2.12: bug #10151 workaround.
Stores lookup parameters + the map from strings to ids
Stores one dependency for the Eisner algorithm Indexes for head and mod start at 1 for the first word in the sentence; 0 is reserved for root
Label information for a dual task that classifies pairs of words (modifier and head) Note: offsets for modifier and head start at 0.
This layer takes a sequence of words and produces a sequence of Expression that stores the words' full embeddings
First layer that occurs in a sequence modeling architecture: goes from words to Expressions
Intermediate layer in a sequence modeling architecture: goes from ExpressionVector to ExpressionVector
A sequence of layers that implements a complete NN architecture for sequence modeling
Multi-task learning (MeTaL) for sequence modeling Designed to model any sequence task (e.g., POS tagging, NER), and SRL
Indexes for pairs of words (modifier and head) Note: offsets for modifier and head start at 0.
Indexes for pairs of words (modifier and head) Note: offsets for modifier and head start at 0. "root" heads have index -1
This layer applies a biLSTM over the sequence of Expressions produced by a previous layer
Manages the tasks in LstmCrfMtl
Some really basic vector math that happens outside of DyNet
Converts the standard CoNLLU syntactic dependency format to Metal
Converts Robert's CoNLLY format (for syntactic dependencies, from his LREC 2020 paper) to Metal
Implements the ConstEmbeddings as a thin wrapper around WordEmbeddingMap with additional functionality to produce embeddings as DyNet Expressions
Evaluates the Eisner algorithm as an unlabeled parsing algorithm
Averages the parameter weights from multiple DyNet model files
Diffs 2 DyNet models Necessary to
Scores the labels assigned to a sequence of words Unlike the CoNLL-2003 scorer, this scorer operates over individual tokens rather than entity spans
Utility methods used by DyNet applications
Label information for a dual task that classifies pairs of words (modifier and head) Note: offsets for modifier and head start at 0. "root" heads have index -1