Awesome

EXEMPLAR

EXEMPLAR is an open relation extraction system originating from a research project at the University of Alberta. Relation extraction is the task of, given a text corpus, identifying relations (e.g., acquisition, spouse, employment) among named entities (e.g., people, organizations). While traditional systems are limited to the relations predetermined by the user, open relation extraction systems like EXEMPLAR are able to identify instances of any relation described in the text.

EXEMPLAR takes text files as input and extracts relations with two or more arguments. For instance, consider the following sentence:

NFL approves Falcons' new stadium in Atlanta. 

Given this sentence, EXEMPLAR extracts an instance of the relation "approve new stadium" whose arguments are "NFL", "Falcons" and "Atlanta".

Relation: approve new stadium
    SUBJ: NFL
    POBJ-OF: Falcons
	POBJ-IN: Atlanta

The role of an argument can be one of the following: SUBJ (subject), DOBJ (direct object) and POBJ (prepositional object). We often append the preposition of a POBJ argument to its role (e.g., "POBJ-IN" for preposition "in"). EXEMPLAR uses heuristics to choose a preposition for a POBJ argument whose preposition is implicit. This is the case for "Falcons" in the above example.

People

• Filipe Mesquita
• Jordan Schmidek
• Denilson Barbosa

Building

Download all dependencies:

$ sh dependencies.sh 

Compile and build jar with all dependencies:

$ sh build.sh 

Running

$ sh exemplar.sh [options] <input> <output>
   -b,--benchmark <arg>   expects input to be a benchmark file (arg = binary | nary)
   -h,--help              shows this message
   -p,--parser <arg>      defines which parser to use (arg = stanford | malt)

• input: path to the document file or directory containing the document files. The tool will recursively look for .txt files in subdirectories.
• output: path to the file where the triples will be stored.
• benchmark option: assumes the input file is formatted according to the ground truth files of our benchmarks for open relation extraction (click to download). These benchmarks are discussed in the paper entitled "Effectiveness and Efficiency of Open Relation Extraction" (see reference in the 'Citing' section). Valid values for this option are 'binary' (for binary relation extraction) and 'nary' (for nary relation extraction). If this option is not provided, the system assumes the input file contains plain text.
• parser option: defines which dependency parser to use. Valid values are 'stanford' and 'malt'. If this option is not provided, the system will use the Malt parser.

Sample Output

The output file contains one relation per line. Fields are separated by a tab in the following order: Subjects, Relation, Objects, Normalized Relation and Sentence. This is the output for our example:

SUBJ:NFL#ORG <tab> approves new stadium <tab> POBJ-OF:Falcons#ORG,,POBJ-IN:Atlanta <tab> approve new stadium <tab> NFL approves Falcons ' new stadium in Atlanta .

The suffix in each argument corresponds to its type. Possible types are person (PER), organization (ORG), location (LOC) and miscellaneous (MISC). Subjects and objects are separated by double comma (",,"), if more than one exists.

Libraries

The main libraries used in this tool are:

• Stanford Parser: tokenization, lemmatization, part-of-speech tagging, named entity recognition and dependency parsing.
• Malt Parser: dependency parsing.

Citing

If you use this code in your research, please acknowledge that by citing:

@INPROCEEDINGS { mesquita-schmidek-barbosa:2013:EMNLP, 
	AUTHOR = { Filipe Mesquita and Jordan Schmidek and Denilson Barbosa }, 
	BOOKTITLE = { Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing }, 
	MONTH = { October }, PAGES = { 447--457 }, 
	PUBLISHER = { Association for Computational Linguistics }, 
	TITLE = { Effectiveness and Efficiency of Open Relation Extraction }, 
	PDF = { http://www.aclweb.org/anthology/D13-1043 }, 
	YEAR = { 2013 }
} 

Acknowledgements

This work was primarily funded by the NSERC Business Intelligence Network (BIN).