Komenti

Komenti is a tool for semantic query, annotation, and analysis of text using ontologies.

It enables querying multiple ontologies with complex class descriptions using AberOWL. These can be used to build a vocabulary for text annotation, including new methods for synonym and label expansion. Annotation is performed using Stanford CoreNLP, and include novel methods for the detection and disambiguation of concept negation and uncertainty. Annotations of text corpora can be used for analysis, within or without Komenti. These components are in development, but currently include summarisation of the co-ocurrence of groups of concepts across text, and use of annotations to suggest description logic axioms for classes. These more complex uses can be described by series of parameters to be passed to the tool in the form of a serialised 'roster,' defining a natural language processing pipeline.

We are working on papers discussing the novel components. I will post them here:

• Komenti: A semantic text mining framework
  • A general description of the tool and a use case with a practical application. Please cite this paper if you use Komenti.
• A fast, accurate, and generalisable heuristic-based negation detection algorithm for clinical text
  • When you use komenti annotate, annotations will automatically be marked for negation and uncertainty using the algorithm described.
• Improved characterisation of clinical text through ontology-based vocabulary expansion
  • You can use this feature by passing the --expand-synonyms flag to the query command.
• Exploring Binary Relations for Ontology Extension and Improved Adaptation to Clinical Text
  • These features are available in the --extract-triples mode of annotate, and in the ontologise command. This is not currently documented, but documentation + a guide is planned for the near future.

You can also find some guides on how to use Komenti in different ways here.

Installation

You can find the latest stable-ish release here: https://github.com/reality/komenti/releases/tag/0.2.0-SNAPSHOT-4

You can add the bin/ directory to your PATH, to be able to use it easily from anywhere. It should also work on Windows, but I haven't tested that.

See here for more information.

Query

Semantic

Get classes and labels satisfying a complex class description using Manchester OWL Syntax.

komenti query -q "'part of' some 'apoptotic process'" -o GO --out labels.txt

Class list

komenti query -c toxicity,asbestos -o ENM --out labels.txt

All classes

You can get all classes in an ontology by running a subclass query on owl#Thing:

komenti query -q "<http://www.w3.org/2002/07/owl#Thing>" --ontology HP

Class mode

If you're interested in querying the ontology classes rather than in creating a text-mining vocabulary, you can run query with --class-mode, which will return one entry per IRI matching the query, with the 'first' label. Note: it currently does not care to identify preferred label by annotation property. For example:

komenti query -q "'Phenotypic abnormality'" --class-mode --direct --ontology HP

Parameters

• The labels can be extended by the power of lemmatisation, by passing --lemmatise
• Synonyms can be expanded used name and semantic matching over AberOWL by passing --expand-synonyms
• --query-type allows you to run either subclass, equivalent, subeq, or superclass queries (the default is subeq)
• You can pass --direct to only retrieve direct (non-transitive) super/subclasses pertaining to your query.
• --class-mode returns only one entry per matching IRI.
• --object-properties allows you to query object properties. If no -q is given, all object properties are returned.
• --override-group will override the group in the output. By default, the value of the group column is the query.

Get Abstracts

Get abstracts from EBI PMCSearch matching any class label (using, as input, the output of the 'query' sub-command).

komenti get_abstracts -l labels.txt --out abstracts/

Parameters

Queries can be grouped by the query used to receive them (the third column in the labels file):

komenti get_abstracts -l labels.txt --group-by-query --out abstracts/

Queries can be conjunctivised (articles must match at least one of every query group):

komenti get_abstracts -l labels.txt --group-by-query --conjunction --out abstracts/

Annotate

Annotate text using labels using Stanford CoreNLP.

komenti annotate -l labels.txt -t text/ --out annotations.txt

Parameters

• -t/--text can be a file or a directory. The files can be text files, or PDF files (whose text will automatically be extracted)
• --family-modifier will add an additional modifier tag for each sentence, indicating whether the sentence mentions a family member (that is, it includes one of the words in the family word list).
• --per-line Annotate each line of each file seperately. This is useful for field-based data, which doesn't have clear line boundaries.
• --disable-modifiers Don't evaluate the annotations for modifiers. These can be added to an annotation file later, using the add_modifiers command.
• --file-list Instead of using --text, you can pass a text file that contains a list of files and directories to annotate, one on each line.
• --group-directory-files will use the name of the parent directory as the document id in the annotations output, instead of the filenames themselves

Summarise Entity Pairs

Summarise the co-occurence of two groups of concepts that have been annotated in a text corpus.

komenti summarise_entity_pairs -l labels.txt -a annotation.txt -c asbestos,toxicity

Diagnose Documents

This analysis tool takes an annotation file as input. It assumes that each annotation file describes one entity, and then for each distinct concept annotated in that file, it decides its overall status with respect to that concept. For example, is hypertension, overall, negated in this document? Or uncertain? If there are separate family flags, then these will have their own separate decision, e.g. a patient may have family history of HCM, but not HCM themselves.

komenti diagnose -a annotations.txt --out diagnoses.txt

The output will be tabular, describing the status for each 'target.' The only targets currently implemented are 'self' and 'family'. The data describes triples, but is formatted in this way to be easier to read (for example, including both the IRI and first label of the concept, and not creating a new predicate to assign the target to an assertion). If the concept is not mentioned, it will not be included.

id	concept iri	concept label	target	status
0001.txt	I50	hypertension	self	negated
0001.txt	I50	hypertension	family	affirmed

Parameters

• --by-group Instead of doing a diagnosis per IRI, a diagnosis per-group will be done. The 'label' column of the output, in this case, will be null.

Generating and Running Rosters

Rosters are files that determine the parameters for series of commands in Komenti. Using them, we can create a specification for a pipeline that runs many Komenti commands, using any outputs in subsequent commands. Here are two examples, the first a general annotation pipeline, the second specifically for examining concept co-occurence:

komenti gen_roster --with-abstracts-download --query "toxicity" --ontology ENM --out roster.json

komenti gen_roster --mine-relationship -c asbestos,toxicity -o ENM --out relationship_roster.json

komenti gen_roster --suggest-axiom --with-metadata-download -c 'nanoparticle' --ontology ENM --entity nanoparticle,nanocage,nanocell,nanosphere,nanohorn,nanorod,nanotube,nanoshell,'quantum dot' --default-entity nanoparticle --quality 'chemical substance','environmental material' --default-relation has_component_part --out enm_roster.json

The rosters can be executed with the following command:

komenti auto -r roster.json

Tips:

• Large vocabularies can have some formatting problems, due to some minor bugs, leading to failure to parse the file at annotate time. It will usually tell you where these are, and it will involve adding or removing some backslashes. I will get around to it. There can also be blank lines sometimes, which must be removed.
• The AberOWL API may become upset if you run too many queries, particularly when trying to --expand-synonyms on all classes from large ontologies. Try reducing the number of threads, and trying again later (there seems to be a throttling thing going on).
• When one argument isn't parsed correctly, the other ones won't work. If it seems to be ignoring arguments passed, check that all your arguments pass. This needs some better error checking code (sorry)