Naisc has a number of ways that it can be configured through the use of Json configuration files, this document describes all the configuration parameters that are available for Naisc. This document is auto-generated based on the annotations in the codebase.
The configuration file consists of the following sections:
blockinglensestextFeaturesgraphFeaturesscorersmatchersrescaler- Other global properties
Each of these is described in more detail in the following section. Every component is specified by an
object with the property name to indicate the component to be used.
Blocking strategies occur in the blocking section of the configuration. There is only a single blocking
strategy so value of blocking is an object with a name property.
The smart, automatic matching strategy that builds on the analysis of the datasets to find potential matches. This setting should be used most of the time
Name: blocking.Automatic
maxMatches: The maximum number of candidates to generate per entity (int)ngrams: The character n-gram to use in matching (Default value: 3) (int)
This blocking strategy matches all possible URIs between the two datasets. It has no configuration parameters.
Name: blocking.All
No parameters
Match according to the identifier. This is used in the case of a dataset where the linking is already known (by the URI) and the goal is to find the semantic similarity. When using this setting pre-linking should be disabled
Name: blocking.IDMatch
method: The method to match (Default value: "endOfPath") (One of exact|fragment|endOfPath|namespace)leftNamespace: The namespace for matching elements in the left dataset (Default value: "") (String)rightNamespace: The namespace for matching elements in the right dataset (Default value: "") (String)
This setting assumes that there is a matching label that indicates candidates. This can be used for example for dictionary sense linking where the goal is to match senses with the same entry, although note the same behaviour is implemented by the OntoLex linker
Name: blocking.LabelMatch
property: The label property to match on (Default value: "http://www.w3.org/2000/01/rdf-schema#label") (String)rightProperty: The label property in the right datasets (if different from left) (Default value: "") (String)language: The language to match on, as an ISO 639 code (Default value: "en") (String)mode: The mode to match; strict for exact matching, or lenient for partial (Default value: "strict") (One of strict|lenient)lowercase: Whether to lowercase labels before matching (Default value: true) (boolean)
String matching generates a blocking that consists of the most similar entities between the two datasets based on a string label. It can be implemented with either Levenshtein or N-Gram similarity
Name: blocking.ApproximateStringMatching
maxMatches: The maximum number of matches to return per entity (int)property: The property to use to find a text label (Default value: http://www.w3.org/2000/01/rdf-schema#label) (String)rightProperty: The property to use in the right dataset. If this is null or omitted then thepropertyis used for both the left and right dataset (String)queueMax: The maximum size of the queue (sets the default queue size, 0 for no limit, only for Levenshtein) (int)metric: The string similarity metric to use (Default value: ngrams) (One of levenshtein|ngrams)ngrams: The maximum size of character n-gram to use in matching (Default value: 3) (int)lowercase: Use case-insensitive matching (Default value: true) (boolean)type: Type of the element. If set all matched elements are of rdf:type with this URI (String)
Used when the blocking is already known. This blocker simply loads a blocking from a file and returns it.
Name: blocking.Predefined
links: The path to the file containing the links to produce (String)
This is used to create a monolingual word sense alignment between two dictionaries in the OntoLex-Lemon format
Name: blocking.OntoLex
No parameters
Name: blocking.Command
command: The command to run it should have to slots $SPARQL_LEFT and $SPARQL_RIGHT for the URL of the left and right SPARQL endpoint (String)
This blocking strategy uses the graph distance based on a number of pre-linked elements. This means that this blocker first looks for a set of elements where there is a value shared by exactly two elements in the left and right dataset, and then returns as candidates all elements that are within n hops in the graph from one of these pre-links
Name: blocking.Path
maxMatches: The maximum number of nodes to explore in the path method (int)preblockLeftProperty: The property to use in the left side of the pre-blocking (String)preblockRightProperty: The property to use in the right side of the pre-blocking (or empty for same as left) (String)
Lens configuration is given in the lenses section of the configuration. There may be multiple lenses and as
such the lenses parameter takes an array of objects, where each object has a name.
Extract a string from a pair of entities by a single property
Name: lens.Label
property: The property to extract (Default value: ["http://www.w3.org/2000/01/rdf-schema#label"]) (String)rightProperty: The property to extract (Default value: ["http://www.w3.org/2000/01/rdf-schema#label"]) (String)language: The language to extract (Default value: null) (String)id: The unique identifier of this lens (String)
Extract a label from the URI itself by de-camel-casing the final part of the URI string
Name: lens.URI
location: The location of the label in the URL (One of fragment|endOfPath|infer)form: The form (camelCased, under_scored) of the label (One of camelCased|underscored|urlEncoded|smart)separator: The character that separates words in the label (String)
A lens that is implemented by a SPARQL query. The query should return exactly two string literals and should contain the special variables $entity1 and $entity2. For example
PREFIX rdfs: <http://www.w3.org/2000/01/rdf-schema#>
SELECT ?label1 ?label2 WHERE {
$entity1 rdfs:label ?label1 .
$entity2 rdfs:label ?label2 .
}Name: lens.SPARQL
query: The SPARQL query (String)baseURI: A base URI for the query (optional) (String)
Analyse a dataset according to the OntoLex model and extract labels accordingly
Name: lens.OntoLex
dialect: The dialect (namespace) to use (One of ONTOLEX|LEMON|MONNET_LEMON)onlyCanonical: Only use canonical forms or use all forms (boolean)language: The language to extract, null for first language available (String)
Name: lens.Command
command: The command to run, the sparql endpoint for the data will be provided as $SPARQL (String)id: The identifier of this feature extractor (String)
Text features are given in the textFeatures section of the configuration. There may be multiple text features
so the textFeatures parameter takes an array of objects, where each object has a name. In addition, you may
provide a tags parameter to any text feature, which selects the lenses it may use.
Similarity based on bag of words
Name: feature.BagOfWordsSim
method: The similarity method to use (One of jaccard|jaccardExponential)weighting: The weighting value. Near-zero values will penalize low agreement morewhile high values will be nearly binary (double)lowerCase: Whether to lowercase the text before processing (boolean)
Basic language-independent string-based similarity
Name: feature.BasicString
labelChar: Also extract character-level features (boolean)wordWeights: Weight the words according to this file (String)ngramWeights: Weight the character n-grams according to this file (String)features: The features to extract (Default value: null) (List of One of lcs|lc_prefix|lc_suffix|ngram_1|ngram_2|ngram_3|ngram_4|ngram_5|jaccard|dice|containment|senLenRatio|aveWordLenRatio|negation|number|jaroWinkler|levenshtein|mongeElkanJaroWinkler|mongeElkanLevenshtein)lowerCase: Convert all strings to lower case before processing (Default value: true) (boolean)
Check for synonyms in a dictionary
Name: feature.Dictionary
dict: The dictionary to use (tab-separated synonyms, one per line) (String)
Keywords feature measures the Jaccard/Dice overlap of a set of key terms.
Name: feature.KeyWords
keywordsFile: The file containing the key words (String)
Similarity based on word embeddings. This method creates a grid of word similarity
Name: feature.WordEmbeddings
embeddingPath: The path to the embeddings file (String)features: The features to use; values include "fp", "bp", "ham", "max", "max2", "max.5", "max.1", "collp2", "collp10", "Hg" (List of String)saliencyFile: The path to the saliency values (String)stopwords: The stopwords file (if used) (String)
Similarity based on the overlap of synonymous and closely related words according to WordNet
Name: feature.WordNet
wordnetXmlFile: The path to the WordNet file in GWA XML format (String)methods: The methods to use (List of One of SHORTEST_PATH|WU_PALMER|LEAKCOCK_CHODOROW|LI)
Name: feature.Command
command: The command to run (String)id: The identifier (String)
String similarity methods based on those widely-used for the evaluation of machine translation
Name: feature.MachineTranslation
methods: The methods to use (Default value: ["BLEU", "BLEU-2", "chrF", "METEOR", "NIST", "TER"]) (List of One of BLEU|BLEU2|chrF|METEOR|NIST|TER)bleuN: The n-gram to use for BLEU (Default value: 4) (int)bleuN2: The n-gram to use for the second BLEU (Default value: 2) (int)chrFN: The n-gram size for chrF (Default value: 6) (int)chrFbeta: The beat paramater for chrF (Default value: 3) (int)nistN: The n-gram size for NIST (Default value: 4) (int)
Graph features are given in the graphFeatures section of the configuration. There may be multiple graph features
so the graphFeatures parameter takes an array of objects, where each object has a name.
Measures the overlap of two entities by properties that they both have. This is useful if there are properties such as part-of-speech or type that can guide the linking
Name: graph.PropertyOverlap
properties: The set of properties to use for overlap or empty for no properties (Set of String)
Name: graph.Command
command: The command to run, the sparql endpoint for the data will be provided as $SPARQL (String)id: The identifier of this feature extractor (String)
The smart, automatic matching strategy that builds on the analysis of the datasets to find potential matches. This setting should be used most of the time
Name: blocking.Automatic
maxMatches: The maximum number of candidates to generate per entity (int)ngrams: The character n-gram to use in matching (Default value: 3) (int)
The Personalised PageRank metric estimates how close two elements in the two datasets. This method relies on pre-links being constructed between the two datasets. The implementation is based on Lofgren, Peter A., et al. "FAST-PPR: scaling personalized pagerank estimation for large graphs." and details of the parameters are in the paper
Name: graph.PPR
No parameters
Scorers are given in the scorers section of the configuration. There may be multiple scorers (associated with
predicting different properties) so this parameter takes an array of objects, where each object has a name.
The scorer simply averages the weight of the scores generated
Name: scorer.Average
weights: The weights to be applied to the features (double[])property: The property to predict (String)softmax: Apply a soft clipping of average using the sigmoid function (boolean)
This scorer learns and applies an optimal scoring given the features. This is a supervised method and must be trained in advance. It is not robust to changes in the features generated so cannot easily be applied to other datasets
Name: scorer.LibSVM
property: The property to output (String)perFeature: Print analysis of features (boolean)
Name: scorer.Command
command: The command to run. Use $MODEL_PATH to indicate the path to the model. (String)trainCommand: The command to run the trainer. Use $MODEL_PATH to indicate the path to the model. (String)property: The property to output (String)
Robust Adaptive Linear Regression. This scorer is based on linear regression but can produce reasonable results for unseen features (assuming some positive correlation). This works better as a supervised model (although not as well as SVM) but is more robust and effective as an unsupervised method as well
Name: scorer.RAdLR
errorFunction: The error function to use in training (One of KullbackLeibler|SoftkullbackLeibler|FMeasure)
The matcher is given in the matcher section of the configuration. It should be a single object with a name.
Simple matcher that outputs all links over a certain score threshold
Name: matcher.Threshold
threshold: The threshold to accept (double)
A special matcher that implements the Hungarian algorithm (a.k.a. MunkRes) to find a matching that gives the highest score given that no element is linked to more than one element in the other dataset
Name: matcher.UniqueAssignment
threshold: The minimum threshold to accept (double)baseProbability: The probability assigned to non-scored examples (double)
The greedy matcher finds a solution given an arbitrary constraint quickly by always taking the highest scoring link. It may produce poorer results than other methods
Name: matcher.Greedy
constraint: The constraint that the searcher will optimize (Constraint - see 'Constriants' section)threshold: The threshold (minimum value to accept) (double)
Beam search finds a matching according to a generic constraint by keeping a list of top solutions found during the search
Name: matcher.BeamSearch
constraint: The constraint that the searcher will optimize (Constraint - see 'Constriants' section)threshold: The threshold (minimum value to accept) (double)beamSize: The size of beam. Trades the speed and memory usage of the algorithm off with the quality of the solution (int)maxIterations: The maxiumum number of iterations to perform (zero for no limit) (int)
Name: matcher.Command
command: The command to run (String)
Find a matching that satisifies an arbitrary constraint by means of the Monte-Carlo Tree Search algorithm
Name: matcher.MonteCarloTreeSearch
ce: The exploration parameter (expert) (double)maxIterations: The maximum number of iterations to perform (int)constraint: The constraint that the searcher will optimize (Constraint - see 'Constriants' section)
Constraints are elements based to some matchers that restrict the kind of linking Naisc can produce. It should be a single object with a name.
A simple constraint that says that the score must be over a threshold
Name: constraint.ThresholdConstraint
threshold: The minimum threshold to accept (double)
The bijective constraint requires that no more than one link exists for each element on the source and/or target dataset
Name: constraint.Bijective
surjection: The type of constraint: bijective means at most one link on the source and target side, surjective means at most one link on the source side, and inverseSurjective means at most one link on the target side (Default value: bijective) (One of surjective|inverseSurjective|bijective)
Rescalers are still experimental, currently you can only configure to use one of the following methods:
NoScaling: Do not rescale the results of the scorerMinMax: Rescale the results of the scorer so that the highest prediction is 1 and the lowest is 0Percentile: Rescale so that the values correspond to the percentile of values that have this value. e.g., 0.5 means that score is exactly the mode of the dataset
The following further parameters are supported by Naisc:
nThreads: The maximum number of threads to use when aligning (int > 0)includeFeatures: The calculated features will be included in the output alignments (can make the alignment files very large!) (boolean)ignorePreexisting: If there are any links between the datasets already they will be discarded and Naisc will only infer new links (boolean)noPrematching: Do not attempt to find unambiguous links and use the full pipeline for every link inference (boolean)