Knowledge Base Question Answering with Topic Units

Authors: Yunshi Lan, Shuohang Wang, Jing Jiang

IJCAI 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments on three commonly used benchmark datasets show that our method consistently works well and outperforms the previous state of the art on two datasets.
Researcher Affiliation Academia Yunshi Lan , Shuohang Wang and Jing Jiang School of Information System, Singapore Management University {yslan.2015, shwang.2014}@phdis.smu.edu.sg, jingjiang@smu.edu.sg
Pseudocode Yes Although these are two separate steps, we jointly learn the parameters in an end-to-end manner. (See Algorithm 1.)
Open Source Code No The paper does not provide a statement about releasing its source code or a link to a code repository.
Open Datasets Yes We evaluate our KBQA method on three benchmark datasets. Web Questions SP (WQSP): This is a dataset that has been widely used for KBQA [Yih et al., 2016]. Complex Web Questions (CWQ): This dataset was introduced by Talmor and Berant [2018]... Simple Questions (SQ): This is another popularly used KBQA dataset, introduced by Bordes et al. [2015]. For WQSP and CWQ, the knowledge base used is the entire Freebase. For SQ, the knowledge base used is a subset of Freebase that comes with the SQ dataset, which is called FB2M.
Dataset Splits Yes Web Questions SP (WQSP): It contains 2848 training questions, 250 development questions and 1639 test questions. Complex Web Questions (CWQ): This dataset was introduced by Talmor and Berant [2018] with the intention to create more complex questions from the Web Questions SQ dataset. CWQ contains 27K, 3K and 3K questions for training, development and test, respectively. Simple Questions (SQ): This is another popularly used KBQA dataset... SQ contains 76K, 11K and 21K for training, development and test, respectively.
Hardware Specification No The paper does not specify any hardware details such as GPU or CPU models used for the experiments.
Software Dependencies No The paper mentions software like S-MART, GloVe, and Adam optimizer, but does not provide specific version numbers for these or other software dependencies.
Experiment Setup Yes We use Adam optimizer with an initial learning rate of 0.001. All hidden vectors are 200-dimensional. All hyper-parameters are turned on the development data. ... We set K to be 3.