Structural Learning with Amortized Inference

Authors: Kai-Wei Chang, Shyam Upadhyay, Gourab Kundu, Dan Roth

AAAI 2015 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We show the efficacy of our method by training a structured SVM using dual coordinate descent for an entity-relation extraction task. Our experiments aim to show the following properties: (1) the amortization techniques significantly reduce the number of inference solver calls, and hence the training time is reduced; (2) with an adaptation on ϵ, the model converges to the same objective function value; (3) with the help of pre-cached examples, we achieve higher amortization ratio and further speed up the training process. Table 1: Ratio of inference engine calls, inference time, inference speedup, final negative dual objective function value ( D(α) in Eq. (6)) and the test performance of each method.
Researcher Affiliation Academia Kai-Wei Chang, Shyam Upadhyay, Gourab Kundu, and Dan Roth Department of Computer Science University of Illinois at Urbana-Champaign {kchang10,upadhya3,kundu2,danr}@illinois.edu
Pseudocode Yes Algorithm 1 A Dual Coordinate Descent Method with Amortized Inference
Open Source Code No The paper states: 'We implemented our algorithms based on a publicly available Structured SVM package4 in JAVA' and provides a footnote link to 'http://cogcomp.cs.illinois.edu/page/software view/JLIS'. This link points to a general software page for a package they used, not specifically the source code for the methodology described in this paper.
Open Datasets Yes We use the annotated corpus from Roth and Yih (2007), which consists of 5,925 sentences. ... We use a dataset, scene3, with 6 labels and 1,211 instances to demonstrate the performance of our method. ... 3Available at http://www.csie.ntu.edu.tw/ cjlin/libsvmtools/ datasets
Dataset Splits Yes performing a 5-fold cross validation for picking the best regularization parameter C for structured SVM on the entity-relation recognition task.
Hardware Specification Yes We implemented our algorithms based on a publicly available Structured SVM package4 in JAVA and conducted experiments on a machine with Xeon E5-2440 processors.
Software Dependencies No The paper mentions 'JAVA' and 'Gurobi' but does not provide specific version numbers for these or any other software dependencies.
Experiment Setup Yes Unless otherwise stated, we show the performance of training Structured SVM model with C = 0.1 with stopping condition δ = 0.1. For Perceptron, we used an averaged Perceptron implementation in JAVA and set the max number of iterations to 100. ... We choose C from {0.01, 0.5 0.1, 0.5, 1}, running 5 folds for each, for a total of 25 runs. δ is set to 0.1 for all folds.