/clontmpV2_Predicting_DBpedia_types

Primary LanguageRMIT LicenseMIT

Predicting DBpedia types

This project contains the code for generating new types on the DBpedia dataset, exploiting its class hierarchy.

The main script (main_predicting_DBpedia_types_P_M_E.R) is the entry point for executing the different experiments avilable. Check out http://es-ta.linkeddata.es/ to explore details and download data to test experiments.

The current version works with the DBpedia ontology, versions 3.9, 2014 and 2016-10, for the English and Spanish datasets.

Getting Started

Clone or download the source code into your machine. Check next section to find libraries and software required and how to install them

Prerequisites

Installing R packages

The use of last stable versions is recommended, but here you point to both, last stable and what the ones we used.

  • Main common libraries and dependencies
#http://docs.h2o.ai/h2o/latest-stable/h2o-docs/faq/r.html
if (! ("methods" %in% rownames(installed.packages()))) { install.packages("methods") }
if (! ("statmod" %in% rownames(installed.packages()))) { install.packages("statmod") }
if (! ("stats" %in% rownames(installed.packages()))) { install.packages("stats") }
if (! ("graphics" %in% rownames(installed.packages()))) { install.packages("graphics") }
if (! ("RCurl" %in% rownames(installed.packages()))) { install.packages("RCurl") }
if (! ("jsonlite" %in% rownames(installed.packages()))) { install.packages("jsonlite") }
if (! ("tools" %in% rownames(installed.packages()))) { install.packages("tools") }
if (! ("utils" %in% rownames(installed.packages()))) { install.packages("utils") }
if (! ("optparse" %in% rownames(installed.packages()))) { install.packages("optparse") }
  • Used ML versions
url_h2o <- "https://cran.r-project.org/src/contrib/Archive/h2o/h2o_3.16.0.1.tar.gz"
install.packages(url_h2o, repos=NULL, type="source")

url_c50 <- "https://cran.r-project.org/src/contrib/Archive/C50/C50_0.1.0-24.tar.gz"
install.packages(url_c50, repos=NULL, type="source")
  • Last ML versions
install.packages("h2o", type="source", repos=(c("http://h2o-release.s3.amazonaws.com/h2o/latest_stable_R")))

if (! ("C50" %in% rownames(installed.packages()))) { install.packages("C50") }

Running experiments

Navigate to the source folder and execute the main script with -h or --help to see options. Use RScript command if you are under Windows. (mind PATH configuration). Every directory that appears as argument is relative (not absolute), and it always should take place under project directory.

$ ./main_predicting_DBpedia_types_P_M_E.R --help
Usage: ./main_predicting_DBpedia_types_P_M_E.R <options>
Description:
                           This software provides a complete Data Mining workflow for inferring new DBpedia types on resources which are found as objects.

Options:
        -a CHARACTER, --approach=CHARACTER
                approach selected. <global_ap1 | multilevel_ap2 | cascade_ap3>.
              global_ap1     - First approach, it learns from most specific type from each resource
              multilevel_ap2 - Second approach, Local Classifiers per Level with binary decisions per level aimed to solve partial depth issue
              cascade_ap3    - Third approach, Local Classifiers per Level with binary decisions per level and cascade process aimed to solve partial depth issue and reduce hierarchy inconcistencies

        -l CHARACTER, --algorithm=CHARACTER
                algorithm used for the approach selected. <NB | C5.0 | DL | RF>.
              NB   - Naïve Bayes. Only in global approach
              C5.0 - Only in multilevel approach
              DL   - Deep Learning
              RF   - Random Forest

        -c LOGICAL, --cross_validation=LOGICAL
                decide whether the divide process should be a simple split (train/test) or should be a cross_validation. <TRUE | FALSE>

        -t INTEGER, --test_ingoingCondition=INTEGER
                minimum amount of properties per resource reserved for tests. Only available on simple split (-c FALSE)

        -n INTEGER, --number_casesOrFolds=INTEGER
                the number of cases or the folds used depending on -c condition option

        -d CHARACTER, --domain=CHARACTER
                how starts URI's DBpedia resources. <EN | ES>

        -v CHARACTER, --versionOntology=CHARACTER
                DBpedia ontology version. <39 | 2014 | 201610>

        -m CHARACTER, --mapping_based_properties=CHARACTER
                path to input mapping_based_properties.ttl. In modern DBpedia versions it may be mapping_based_objects.ttl, it works well so.

        -i CHARACTER, --instance_types=CHARACTER
                path to input instance_types.ttl In modern DBpedia versions it may be splitted in 2 files, transitive and 'ending' (leaves) types.
              This main program assumes that types file are completed, so in case of divided files, should be merged previously

        -o CHARACTER, --pathOut=CHARACTER
                path to output files. Directory should exist previously

        -e CHARACTER, --pathIntermediateData=CHARACTER
                path to intermediate data generated in process, as train/test partition sets.

        -f CHARACTER, --fileOut=CHARACTER
                files' output identifier or name to track files about same experiment [default= output]

        -s INTEGER, --seed=INTEGER
                random number generator seed for algorithms that are dependent on randomization [default= 1234]

        -x CHARACTER, --executionMode=CHARACTER
                this main is divided in 3 modules: preprocesing, modeling, evaluating. Flag -x or --executionMode specify where this main script should do. This is useful, for instance, to generate the same train/validation data once and execute several approaches and/or algorithms. <P_M_E | P | M_E >

        -h, --help
                Show this help message and exit


                                      Note: each directory used in arguments as -e or -f should end with slash ('/')
                                      Examples:
                             ./main_predicting_DBpedia_types_P_M_E.R -a global_ap1 -l DL -c FALSE -t 1 -n 2500 -d ES -v 201610 -m ~/inputData/mappingbased_objects_uncleaned_es.ttl -i ~/inputData/instance_types_completo_es.ttl -o ~/outputData/ -f ejecucion1_app1_DL_test1 -s 1234 -x P_M_E

                            In this case, The program will use first approach (global) with a deep learning algoritm, with one train/validate split where 2500 validation cases have at least 1 ingoing property. This case is used on EsDBpedia with the 2016-10 ontology. After that, we specify both input data (properties and types) followed by output path, and files identifier. All processes will use '1234' as random seed. The -x option will indicate a full workflow (preprocesing, modeling and evaluating)

There are a couple of bash scripts that can help to run several experiments together sequentially. Please, check them and adapt them to your own experiments before run them. Some example executions:

./lanzador_lanzaMains_V5.sh "ES" "201610" 2500 inputData/es201610/instance_types_completo_es.ttl inputData/es201610/mappingbased_objects_uncleaned_es.ttl

./lanzador_lanzaMains_V5.sh "EN" "201610" 10000 inputData/en201610/instance_types_completo_en.ttl inputData/en201610/mappingbased_objects_uncleaned_en.ttl

About the authors

Acknowledgments

This work was partially funded by projects RTC-2016-4952-7, TIN2013-46238-C4-2-R and TIN2016-78011-C4-4-R, from the Spanish State Investigation Agency of the MINECO and FEDER Funds