ConceptNet 5.5: An Open Multilingual Graph of General Knowledge

Authors: Robyn Speer, Joshua Chin, Catherine Havasi

AAAI 2017 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We demonstrate this with state-of-the-art results on intrinsic evaluations of word relatedness that translate into improvements on applications of word vectors, including solving SAT-style analogies.
Researcher Affiliation Collaboration Robyn Speer Luminoso Technologies, Inc. ... Joshua Chin Union College ... Catherine Havasi Luminoso Technologies, Inc.
Pseudocode No No pseudocode or algorithm blocks are present in the paper.
Open Source Code Yes The code and documentation of Concept Net 5.5 can be found on Git Hub at https://github.com/commonsense/ conceptnet5, and the knowledge graph can be browsed at http://conceptnet.io.
Open Datasets Yes We focus on MEN-3000 (Bruni, Tran, and Baroni 2014), a large crowd-sourced ranking of common words; RW (Luong, Socher, and Manning 2013), a ranking of rare words; Word Sim-353 (Finkelstein et al. 2001), a smaller evaluation that has been used as a benchmark for many methods; and MTurk-771 (Halawi et al. 2012), another crowd-sourced evaluation of a variety of words. AND On a corpus of SAT-style analogy questions (Turney 2006).
Dataset Splits Yes MEN-3000 is already divided into a 2000-item development set and a 1000-item test set. We use the results from the test set as the final results. AND RW has no standard dev/test breakdown. We sampled 2/3 of its items as a development set and held out the other 1/3 (every third line of the file, starting with the third). AND We used all of Word Sim-353 in development. AND The grid search is performed on odd-numbered questions, holding out the even-numbered questions as a test set.
Hardware Specification No No specific hardware details (e.g., GPU/CPU models, memory specifications) are provided for the experiments.
Software Dependencies No The paper mentions Snakemake and Docker Compose but does not specify version numbers for other key software dependencies or libraries used in the experiments.
Experiment Setup Yes The weights found for Concept Net Numberbatch 16.09 were w1 = 0.2 and w2 = 0.6. AND Our preliminary attempt to apply Concept Net Numberbatch to the Story Cloze Test is to use a very simple bag-of-vectors model, by averaging the embeddings of the words in the sentence and choosing the ending whose average is closest.