The ‘RmrsDmrsComparison’ tool operates on DMRS syntactic representation of sentences in XML format and stores them in memory using efficient data structures. These files contain many sentences each with many associated parses. The tool has five different usages triggered by different command line arguments.
The packing and unpacking tool output the internal representation to an XML file in packed and unpacked DMRS representations respectively. The structural comparison function takes two sets of DMRS sentences and identifies which sentences are the same, different or partially match. Structural similarity identification compares two parses for common nodes and links and finally the duplicate node and link identification tool finds duplicate nodes and links within a parse.
Source code and a ready to run .jar file can be downloaded from the google code repository.
Complete implementation details and framework explanations can be also be found under the repository download section in thesis.pdf.
A brief explanation of functionality will be given here for convenience.
The tool has 5 different functions triggered by command line tags. A brief reminder of these usages can be found by running the tool with no arguments.
USAGE 1: -compare [options] DMRS1in DMRS2in
USAGE 2: -pack [options] DMRSin out
USAGE 3: -unpack [options] DMRSin out
USAGE 4: -parseSim [options] DMRS1in DMRS2in pairFile
USAGE 5: -duplicate [options] DMRS1in
Each of these usages are defined below.
The input format is flexible. The command line input source paths can point to a single file or a directory of related files. These files can be either packed or unpacked in either the the standard or alternative DMRS formats. Files can also optionally be gzipped. A file ending in .gz is assumed to be gzipped.
The XML specifications for the original and alternative versions of the unpacked and packed DMRS representations can be downloaded from the downloads section of the code hosting.
The tool accepts several options that can be passed with the command line arguments.
-o out Output File
-gzip Output files using GZip
-s Do not use sortinfo in node comparison
-sh Output to shell ON
-alt Output alternative XML format
-o and its argument 'out' specify the file where results will be output.
-gzip forces all output files to be gxzipped.
-s ignores the 'sortinfo' fields in the DMRS node structures during comparison. This field cannot be set when packing or unpacking.
-sh prints results to the shell as well as to any specified file.
-alt changes the DMRS output format to the alternative format. This is only applicable for functions that output XML files.
The packing function converts any input to a packed representation and outputs this as unpacked XML
The unpacking functions converts any input to an unpacked representation and outputs this as packed XML
The comparison function reads two sets of DMRS representations (created using the same test set but possibly different grammars) and compares the DMRS semantic structure.
Two sentences are considered equal if all parses in the sentence are equal to at least one parse in the opposing sentence. This is similar to a surjective relation except allows for spurious ambiguity (a single sentence contains two or more equal parses). Also, if two sentences are not equal they may be considered 'partially equal' (some but not all parses map) or 'not equal' (no parse in either sentence matches a parse in the other). Two parses are considered equal if there is a bijective equality mapping between their respective sets of nodes and links. In other words if two parses contain the same number of links and nodes, and every node and link in the one parse matches a node or link in the second parse then they are considered to be equal. The inclusion of ‘sortinfo’ is optional.
The output format of this function is as follows :
[sentence no.] [match status] – [number of parses in A] & [number of parses in A] parses – [number of parses that match] - [a list of matches with the format ‘(ID number from A : ID numbers in B that match)’]
The similarity algorithm identifies node and link commonalities and outputs them to a file. Nodes and links are considered equal if they are equal in all their fields except for ID numbers. The inclusion of 'sortinfo' is optional.
The output format of this function is a dmrs-pair list. This is very similar to the standard unpacked DMRS format except nodes contain two ID numbers, one for each of the nodes that match, and links contain two sets of node ID links, representing the nodes connected by the respective matching links. The output XML specification can be found on the source code hosting downloads page.
In the case where multiple nodes or links in one set match one or multiple nodes or links in the other, then every combination of nodes and links are output.
The duplication tool analyses a single set of DMRS outputs for duplicate nodes and links. When a duplicate expression is found the sentence ID and parse ID it belongs to are output. The inclusion of 'sortinfo' is optional.
The output is a simple ID pair list. If multiple duplication occurs in a file the ID numbers will be output several times.