Faithful Embeddings for Knowledge Base Queries

Authors: Haitian Sun, Andrew Arnold, Tania Bedrax Weiss, Fernando Pereira, William W. Cohen

NeurIPS 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate Em QL first intrinsically for its ability to model set expressions [10], and then extrinsically as the reasoning component in two multi-hop KB question answering benchmarks (KBQA). The numbers are shown in Table 2. Following the Query2Box paper [17] we use d = 400 for their model and report Hits@3 (see Supplementary Materials D for other metrics). To see if larger embeddings would improve Q2B s performance on entailment tasks, we increased the dimension size to d = 2000, and observed a decrease in performance, relative to the tuned value d = 400 [17].
Researcher Affiliation Industry Haitian Sun Andrew O. Arnold Tania Bedrax-Weiss Fernando Pereira William W. Cohen Google Research {haitiansun,tbedrax,pereira,wcohen}@google.com AWS AI anarnld@amazon.com
Pseudocode No The paper describes algorithmic steps in prose, but does not include any clearly labeled 'Pseudocode' or 'Algorithm' blocks.
Open Source Code Yes Code available at https://github.com/google-research/language
Open Datasets Yes The Meta QA datasets [29] contain multi-hop questions in the movie domain that are answerable using the Wiki Movies KB [13]... The Web Questions SP model. This dataset [27] contains 4,737 natural language questions generated from Freebase.
Dataset Splits Yes To evaluate performance for QE, Ren et al. first hold out some triples from the KB for validation and test, and take the remaining triples as the training KB.
Hardware Specification No Here d = 2000 was the largest value of d supported by our GPUs. (Mentions 'GPUs' but no specific models or other hardware details).
Software Dependencies No No specific software dependencies with version numbers (e.g., 'Python 3.8', 'PyTorch 1.9', 'CUDA 11.1') are provided.
Experiment Setup Yes For Em QL, we use d = 64, k = 1000, NW = 2000 and ND = 20 throughout.