First Look: Text Analytics with InterSystems IRIS
This quick start guide introduces you to InterSystems IRIS™ support for Natural Language Processing (NLP) text analytics, which provides semantic analysis of unstructured text data in a variety of natural languages. This enables you to discover useful information about the contents of a large number of text documents without any prior knowledge of the contents of the texts. This quick start guide covers the following topics:
This First Look guide presents an introduction to InterSystems IRIS Natural Language Processing, and walks through some initial tasks associated with indexing text data for semantic text analysis. Once you’ve completed this exploration, you will have indexed a group of texts and performed analysis determining the most common entities in those texts, metrics about those entities, various kinds of associations found between entities, and viewing the appearances of an entity in the source texts. These activities are designed to use only the default settings and features, so that you can acquaint yourself with the fundamentals of NLP text analysis. For the full documentation on Text Analytics, see the InterSystems IRIS Natural Language Processing (NLP) Guide
A related, but separate, tool for handling unstructured texts is InterSystems IRIS SQL Search
. SQL Search allows you to search
for these same entities, as well as single words, regular expressions and other constructs in multiple texts. Inherently, a search solution presupposes that you know what you are looking for. NLP text analytics is designed to help you discover content and connections between content entities without necessarily starting from an idea to look for.
Why NLP Text Analytics Is Important
Increasingly, organizations are amassing larger and larger quantities of unstructured text data, far in excess of their ability to read or catalog these texts. Frequently, an organization may have little or no idea what the contents of these text documents are. Conventional top-down text analysis based on pure search
technologies makes assumptions about the contents of these texts, which may miss important content.
InterSystems IRIS Natural Language Processing (NLP) allows you to perform text analysis on these texts without any upfront knowledge of the subject matter. It does this by applying language-specific rules that identify semantic entities. Because these rules are specific to the language, not the content, NLP can provide insight into the contents of texts without the use of a dictionary or ontology.
How InterSystems IRIS Implements NLP Text Analytics
To prepare texts for NLP analytics you must load those texts into a domain, and then build the domain. Based on its analysis of the texts, NLP builds indices for the domain that NLP can use to rapidly analyze large quantities of text. Texts can be input from a variety of data locations, including SQL tables, text files, strings, globals, and RSS data.
NLP supports the following functionality:
Language models: identifying semantic relationships between words is language-specific. NLP contains semantic rules (language models) for ten natural languages that enable analysis of a text on any subject written in that language. If you specify more than one language, NLP performs automatic language identification by determining the best match between each sentence in each text and the specified languages. NLP analysis does not require the upfront creation or association of dictionaries or ontologies, although you can expand its functionality by adding them.
Entity analysis: NLP operates on semantic groups of one or more words known as entities. Entities are identified as either Concepts (which include nouns and noun phrases) or Relations (which include verbs and prepositions). Commonly, the most relevant entities to consider are Concepts, though it is also possible to analyze Relations. Sentence and word boundaries are always observed. Letter case is ignored.
Path analysis: NLP groups coherent sequences of Concepts and Relations into Paths. A sentence usually consists of a single Path. A Path reveals the connections between entities.
Attributes: NLP flags semantic attributes such as negation, so that you can differentiate text sequences that are positive (evidence of structural damage) from those that are negative (no evidence of structural damage).
Frequency, Spread, and Dominance: These are metrics calculated for an entity. Frequency is the number of times an entity appears in a group of texts. Spread is the number of texts that contain that entity. Dominance is a more nuanced metric generated by factoring in the entity frequency relative to the length of each text, the frequency of other entities having words in common, and other factors. Entities are commonly returned sorted in descending order by these metrics. These metrics provide insight into the content of texts, enabling you to perform deeper analysis of specific entities.
Similar Entities, Related Concepts, and Proximity Profile. Given an entity, these discover other relevant entities. For example, given a short entity, similar entities would include other, longer entities in the domain that contain the same words, thereby being more specific entities than the seed one. Given an entity, related entities are other entities in the same sentence that are associated to the specified entity by a single Relation. Given an entity, the Proximity metric calculates the proximity within paths between the specified entity and other entities.
Dictionaries: you can add optional dictionaries to identify synonyms for an entity.
Summarization: you can use NLP to generate a summary of a text, requesting the summary as a percentage of the whole. For example, a 50% summary would consist of half of the sentences in the original text, with NLP selecting those sentences that are calculated as most relevant to the overall source text.
Trying NLP Text Analytics for Yourself
It’s easy to use InterSystems IRIS Text Analytics. This simple procedure walks you through the basic steps of generating NLP metrics.
You need to have an InterSystems IRIS instance that is up and running and has an active license key. (You can view the licence key from the Management Portal
: select System Administration
All NLP analysis occurs within a domain. You associate multiple texts with a domain. You then build the domain, creating indices that are used by NLP queries.
A domain is created within a namespace. You can create multiple domains within a namespace. You can associate a text with multiple domains.
There are several ways to create, populate, and build a domain. The following example uses the Domain Architect
The starting point for accessing the Domain Architect
is the Management Portal Analytics
option. From there you select the Text Analytics
option. This displays the Domain Architect
All NLP domains exist within a specific namespace. Therefore, you must specify which namespace you wish to use by selecting the Switch
option at the top of any Management Portal
interface page. This displays the list of available analytics namespaces. Select TESTSAMPLES from this list.
From the Domain Architect
button to define a domain. You specify the following domain values (in the specified order):
: The name you assign to a domain must be unique for the current namespace (not just unique within its package class); domain names are not case-sensitive. For this example, specify MyTest
Allow Custom Updates
: this check box enables adding data or dictionaries to this domain manually. For this example, do not check the box.
Within a domain you can define data locations and other model elements for the domain. To add or modify model elements, click on the expansion triangle next to one of the headings. Initially, no expansion occurs. Once you have defined some model elements, clicking the expansion triangle shows the model elements you have defined.
Click the Data Locations
triangle to display the Details tab on the right side of the screen. The Details tab shows five Add Data options. Select Add data from table
This option allows you to specify data stored in an SQL table. In this example we will specify the following fields:
The Domain Architect
page heading is followed by an asterisk (*) if there are unsaved changes to the current domain definition. Click Save
to save your changes.
Compile the Domain by pressing the Compile
Then build the NLP indices for the data sources by pressing the Build
The Domain Explorer
initially displays a list of the most significant concepts in the source texts:
tab displays the Top Concepts
in descending order by frequency. Each listed item is shown with its frequency count (number of occurrences) and spread count (number of sources containing that concept).
For example, the concept pilot
has a frequency of 6206 and a spread of 1085; the concept student pilot
has a frequency of 319 and a spread of 141.
For example, the concept pilot
has a dominance of 351.6008; the concept student pilot
has a dominance of 49.3625.
When you select one of these concepts the other Domain Explorer
listings are displayed:
lists the selected concept, and any other concepts that contain that concept, each with its frequency and spread.
lists other concepts that are related to the selected concept, with the frequency and spread of the instances of those concepts in this related context.
lists other concepts that found in proximity to the selected concept, with calculated proximity score of the instances of those concepts when found in the same sentence as the selected concept.
By selecting a concept in any of these lists, these listings are refreshed based on that concept. Alternatively, you can also type an entity (Concept or Relation) into the Domain Explorer Explore
area and click the Explore!
By using these listings, you can determine what concepts appear in the source documents, how significant they are, and what other concepts are associated with them.
The lower portion of the Domain Explorer
allows you to view how a selected concept appears in the source texts:
tab lists by source all of the sentences that contain the selected concept. The concept is highlighted and red text is used to indicate negation that involves the concept.
tab lists all of the paths that contain the selected concept. The path text is highlighted to show NLP indexing: the selected concept is highlighted in orange, other concepts in the path are highlighted in blue, path-relevant concepts (commonly pronouns) are highlighted in light blue. Relations are shown white. Red text is used to indicate negation that involves the concept.
By clicking the eye
icon, you can display the complete text of the source, with the selected concept highlighted, and red text used to indicate negation.
toggle button displays the complete text of the source with highlighting showing NLP indexing. Because this is the source text, capitalization, punctuation, and non-relevant words are shown; these aspects of the text are not shown in the Paths
option allows you to display a summary of the text. Specify a percentage. The total number of sentences in the text is reduced to that percentage. The sentences that NLP includes in the summary are determined by metrics calculating their significance to the full text.
Often the list of top concepts begins with concepts that are too common or concepts that have little value in discovering useful information. These may be words or phrases that appear in all of the sources (such as accident report or conclusions), general concepts (such as airplane or pilot), or concepts not relevant to your use of the data (such as a list of cities). You can use a blacklist to prevent the display of these concepts. A blacklist only affects the display of concepts in certain query results; it has no effect on NLP indexing of concepts.
In the Domain Architect
click the Open
button and select Samples >>
to open the existing domain Samples.MyTest.
Click the Blacklists
expansion triangle. This displays the Add blacklist
button in the Details tab on the right side of the screen. Click Add blacklist
to display the Name and Entries fields. Accept the default name for the blacklist (Blacklist_1
). In the Entries
box list entries (concepts) one concept per line; entries are not case-sensitive. In this example list the concepts: pilot, student pilot, co-pilot, passenger, instructor, flight instructor, certified flight instructor.
the domain. (You do not need to Build
the domain to add, modify, or remove blacklists).
In the Domain Explorer
click the sunglasses
icon in the upper right corner. This displays a list of the blacklists defined for this domain that you can apply. Select Blacklist_1
. Note that the Top Concepts
listing no longer lists the blacklist concepts.
This example is provided to give you some initial experience with InterSystems IRIS Natural Language Processing. You should not use this example as the basis for developing a real application. To use NLP in a real situation you should fully research the available choices provided by the software, then develop your application to create robust and efficient code.
For More Information about NLP Text Analytics
InterSystems has other resources to help you learn more about NLP Text Analytics: