Joint Morphological Generation and Syntactic Linearization

Authors: Linfeng Song, Yue Zhang, Kai Song, Qun Liu

AAAI 2014 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments show that the joint method significantly outperforms a strong pipelined baseline (by 1.1 BLEU points). It also achieves the best reported result on the Generation Challenge 2011 shared task. We perform experiments on the corpus for the Generation Challenges 2011 Surface Realisation Shared Task
Researcher Affiliation Academia 1Key Laboratory of Intelligent Information Processing Institute of Computing Technology, Chinese Academy of Science 2Singapore University of Design and Technology 3CNGL, School of Computing, Dublin City University 4Natrual Language Processing Lab., Northeastern University
Pseudocode Yes Algorithm 1: Decoding algorithm for linearization. Algorithm 2: The Agenda Initiation Function for the Joint algorithm.
Open Source Code Yes We release our code at https://sourceforge.net/projects/zgen/
Open Datasets Yes We perform experiments on the corpus for the Generation Challenges 2011 Surface Realisation Shared Task1, which provides training and test data for the shallow syntax linearization to morphological generation pipeline. 1http://www.nltg.brighton.ac.uk/research/sr-task/
Dataset Splits Yes For development, we extract 1 out of every 20 sentences from the original training set as the development test set, and take the remaining training data as the development training set. Our development test set contains 1809 sentences and the development training set contains 34400 sentences.
Hardware Specification Yes All the systems were implemented in Python, and all the experiments were performed on an Intel Xeon E5 2660 CPU with Centos 6.3 and Python 2.6.
Software Dependencies Yes All the systems were implemented in Python, and all the experiments were performed on an Intel Xeon E5 2660 CPU with Centos 6.3 and Python 2.6.
Experiment Setup Yes We perform 10 training rounds for the morphological generation system, 20 iterations for the linearization system and 20 iterations for the joint system, according to development experiments. We set the timeout threshold of both the linearization system and the joint system to 8s for all tests