A Semi-Supervised Learning Approach to Why-Question Answering

Authors: Jong-Hoon Oh, Kentaro Torisawa, Chikara Hashimoto, Ryu Iida, Masahiro Tanaka, Julien Kloetzer

AAAI 2016 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Through a series of experiments, we showed that our approach significantly improved the precision of the top answer by 8% over the current state-of-the-art system for Japanese why-QA.
Researcher Affiliation Academia National Institute of Information and Communications Technology (NICT), Kyoto, 619-0289, Japan
Pseudocode Yes The pseudo-code of our semi-supervised learning algorithm is given in Figure 2.
Open Source Code No The paper does not provide a statement or link indicating that the source code for their proposed methodology is publicly available.
Open Datasets Yes For the experiments, we used the same data set as the one used in our previous works (Oh et al. 2012; 2013). This data set is composed of 850 Japanese why-questions and their top-20 answer passages obtained from 600 million Japanese web texts by using the answer-retrieval method of Murata et al. (2007).
Dataset Splits Yes For our experiments, we divided the data set (17,000 question-passage pairs) into training, development, and test data. For our training data, we first selected 7,000 question-passage pairs for 350 questions, which we used only as training data in our previous works (Oh et al. 2012; 2013) and randomly chose 8,000 question-passage pairs for 400 questions from the remainders of the first selection. We equally divided the 2,000 question-answer pairs for 100 questions, which is the remainder after the selection of the training data, into development and test data.
Hardware Specification No The paper does not specify any hardware details (e.g., CPU, GPU models, memory, or cloud instance types) used for running the experiments.
Software Dependencies No The paper mentions software components like "Lucene search engine", "SVMs", "SVMLight", "J.Dep P5", and "Tiny SVM", but it does not provide specific version numbers for these, which is necessary for reproducibility.
Experiment Setup Yes All the combinations of α, β, and K derived from α {0.2, 0.3, 0.4}, β {0.6, 0.7, 0.8}, and K {150, 300, 450} were tested over the development data, and α = 0.3, β = 0.7, and K = 150, which showed the best performance, were used for our final experiments. We also set the maximum iteration number (l in Figure 2) to 40, where the performance converged in the test with these selected parameters on the development data.