LP-SparseMAP: Differentiable Relaxed Optimization for Sparse Structured Prediction
Authors: Vlad Niculae, Andre Martins
ICML 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments in three structured tasks show benefits versus Sparse MAP and Structured SVM. |
| Researcher Affiliation | Collaboration | 1Instituto de Telecomunicações, Lisbon, Portugal 2Unbabel, Lisbon, Portugal 3Instituto Superior Técnico, University of Lisbon, Portugal. |
| Pseudocode | Yes | Algorithm 1 ADMM for LP-Sparse MAP |
| Open Source Code | Yes | Our library, with C++ and python frontends, is available at https://github.com/deep-spin/lp-sparsemap. |
| Open Datasets | Yes | The List Ops dataset (Nangia and Bowman, 2018) is a synthetic collection of bracketed expressions... We use the English language SNLI and Multi NLI datasets (Bowman et al., 2015; Williams et al., 2017), with the same preprocessing and splits as Niculae et al. (2018). ...bibtex and bookmarks benchmark datasets (Katakis et al., 2008). |
| Dataset Splits | No | The paper mentions using 'validation' sets (e.g., 'validation test Acc. F1' in Table 2, 'valid test' in Table 3) and states using 'the same preprocessing and splits as Niculae et al. (2018)' for SNLI and Multi NLI, but does not explicitly state the specific percentages, sample counts, or detailed splitting methodology within this paper. |
| Hardware Specification | No | The paper does not provide specific hardware details (e.g., GPU/CPU models, memory, or cloud instance types) used for running its experiments. |
| Software Dependencies | No | We use Dynet (Neubig et al., 2017)... Our library, with C++ and python frontends... we acknowledge the scientific Python stack (Van Rossum and Drake, 2009; Oliphant, 2006; Walt et al., 2011; Virtanen et al., 2020; Behnel et al., 2011) and the developers of Eigen (Guennebaud et al., 2010). |
| Experiment Setup | No | The paper states 'We use Dynet (Neubig et al., 2017) and list hyperparameter configurations and ranges in App. E.' and 'all our models have 130k parameters (cf. App. E).', but the detailed hyperparameter values and training configurations are deferred to Appendix E, which is not part of the provided text. |