Intent Matching

Overview

Data Model processing logic is defined as a collection of one or more intents. The sections below explain what intent is, how to define it in your model, and how it works.

Intent

The goal of the data model implementation is to take the user input text and match it to a specific user-defined code that will execute for that input. The mechanism that provides this matching is called an intent.

The intent generally refers to the goal that the end-user had in mind when speaking or typing the input utterance. The intent has a declarative part or template written in IDL - Intent Definition Language that strictly defines a particular form the user input. Intent is also bound to a callback method that will be executed when that intent, i.e. its template, is detected as the best match for a given input. A typical data model will have multiple intents defined for each form of the expected user input that model wants to react to.

For example, a data model for banking chatbot or analytics application can have multiple intents for each domain-specific group of input such as opening an account, closing an account, transferring money, getting statements, etc.

Intents can be specific or generic in terms of what input they match. Multiple intents can overlap and NLPCraft will disambiguate such cases to select the intent with the overall best match. In general, the most specific intent match wins.

IDL Syntax

NLPCraft intents are written in Intent Definition Language (IDL). IDL is a relatively straightforward declarative language. For example, here's a simple intent x with two terms a and b:

            intent=x
                term(a)~{# == 'my_elm'}
                term(b)={has(tok_groups, "my_group")}
        

IDL intent defines a match between the parsed user input represented as the collection of tokens, and the user-define callback method. IDL intents are bound to their callbacks via Java annotation and can be located in the same Java annotations or placed in model YAML/JSON file as well as in external *.idl files.

You can review the formal ANTLR4 grammar for IDL, but here are the general properties of IDL:

• IDL has context-free grammar. In simpler terms, all whitespaces outside of string literals are ignored.
• IDL supports Java-style comments, both single line // Comment. as well as multi-line /* Comment. */.
• String literals can use either single quotes ('text') or double quotes ("text") simplifying IDL usage in JSON or Java languages - you don't have to escape double quotes. Both quotes can be escaped in string, i.e. "text with \" quote" or 'text with \' quote'
• Built-in literals true, false and null for boolean and null values.
• Algebraic and logical expression including operator precedence follow standard Java language conventions.
• Both integer and real numeric literals can use underscore '_' character for separation as in 200_000.
• Numeric literals use Java string conversions.
• IDL has only 10 reserved keywords: flow fragment import intent meta options term true false null
• Identifiers and literals can use the same Unicode space as Java.
• IDL provides over 150 built-in functions to aid in intent matching. IDL functions are pure immutable mathematical functions that work on a runtime stack. In other words, they look like Python functions: IDL length(trim(" text ")) vs. OOP-style " text ".trim().length().
• IDL is a lazily evaluated language, i.e. expressions are evaluated only when required during runtime. That means that evaluated left-to-right logical AND and OR operators, for example, skip their right-part expressions if the left expression result is determinative for the overall result - so-called short-circuit evaluation. Some IDL functions like if and or_else also provide the similar short-circuit evaluation.

IDL program consists of intent, fragment, or import statements in any order or combination:

• intent statement

  Intent is defined as one or more terms. Each term is a predicate over a instance of NCToken interface. For an intent to match all of its terms have to evaluate to true. Intent definition can be informally explained using the following full-feature example:

                      intent=xa
                          flow="^(?:login)(^:logout)*$"
                          meta={'enabled': true}
                          term(a)={month >= 6 && !# != "z" && meta_intent('enabled') == true}[1,3]
                          term(b)~{
                              @usrTypes = meta_model('user_types')
  
                              (# == 'order' || # == 'order_cancel') && has_all(@usrTypes, list(1, 2, 3))
                          }
  
                      intent=xb
                          options={
                              'ordered': false,
                              'unused_free_words': true,
                              'unused_sys_toks': true,
                              'unused_usr_toks': false,
                              'allow_stm_only': false
                          }
                          flow=/#flowModelMethod/
                          term(a)=/org.mypackage.MyClass#termMethod/?
                          fragment(frag, {'p1': 25, 'p2': {'a': false}})
                  

  NOTES:

  intent=xa ^{line 1}
  intent=xb ^{line 12}

  xa and xb are the mandatory intent IDs. Intent ID is any arbitrary unique string matching the following lexer template: (UNI_CHAR|UNDERSCORE|LETTER|DOLLAR)+(UNI_CHAR|DOLLAR|LETTER|[0-9]|COLON|MINUS|UNDERSCORE)*

  options={...} ^{line 13}

  Optional. Matching options specified as JSON object. Entire JSON object as well as each individual JSON field is optional. Allows to customize the matching algorithm for this intent:

  Option	Type	Description	Default Value
  ordered	Boolean	Whether or not this intent is ordered. For ordered intent the specified order of terms is important for matching this intent. If intent is unordered its terms can be found in any order in the input text. Note that ordered intent significantly limits the user input it can match. In most cases the ordered intent is only applicable to processing of a formal grammar (like a programming language) and mostly unsuitable for the natural language processing. Note that while the ordered flag affect entire intent and all its terms, you can define the individual term that depends on the position of the token. This, in fact, allows you have a subset of terms that order dependant. See the following IDL functions for details: tok_index() tok_all() tok_count() tok_is_last() tok_is_first() tok_is_before_id() tok_is_before_parent() tok_is_before_group() tok_is_between_ids() tok_is_between_parents() tok_is_between_groups() tok_is_after_id() tok_is_after_parent() tok_is_after_group()	false
  unused_free_words	Boolean	Whether or not free words - that are unused by intent matching - should be ignored (value true) or reject the intent match (value false). Free words are the words in the user input that were not recognized as any user or system token. Typically, for the natural language comprehension it is safe to ignore free words. For the formal grammar, however, this could make the matching logic too loose.	true
  unused_sys_toks	Boolean	Whether or not unused system tokens should be ignored (value true) or reject the intent match (value false). By default, tne unused system tokens are ignored.	true
  unused_usr_toks	Boolean	Whether or not unused user-defined tokens should be ignored (value true) or reject the intent match (value false). By default, tne unused user tokens are not ignored since it is assumed that user would define his or her own tokens on purpose and construct the intent logic appropriate.	false
  allow_stm_only	Boolean	Whether or not the intent can match when all of the matching tokens came from STM. By default, this special case is disabled (value false). However, in specific intents designed for free-form language comprehension scenario, like, for example, SMS messaging - you may want to enable this option.	false

  flow="^(?:login)(^:logout)*$" ^{line 2}
  flow=/#flowModelMethod/ ^{line 20}

  Optional. Dialog flow is a history of previously matched intents to match on. If provided, the intent will first match on the history of the previously matched intents before processing its terms. There are two way to define a match on the dialog flow:

  • Regular Expression

    In this case dialog flow specification is a string with the standard Java regular expression. The history of previously matched intents is presented as a space separated string of intent IDs that were selected as the best match during the current conversation, in the chronological order with the most recent matched intent ID being the first element in the string. Dialog flow regular expression will be matched against that string representing intent IDs.

    In the line 2, the ^(?:login)(^:logout)*$ dialog flow regular expression defines that intent should only match when the immediate previous intent was login and no logout intents are in the history. If the history is "login order order" - this intent will match. However, for "login logout" or "order login" history this dialog flow will not match.

  • User-Defined Callback

    In this case the dialog flow specification is defined as a callback in a form /x.y.z.Cass#method/, where x.y.z.Class should be a fully qualified name of the class where callback is defined, and method must be the name of the callback method. This method should take one parameter of type java.util.List[NCDialogFlowItem] and return boolean result.

    Class name is optional in which case the model class will be used by default. Note that if the custom class is in fact specified, the instance of this class will be created for each dialog flow test. This class must have a no-arg constructor to instantiate via standard Java reflection and its creation must be as light as possible to avoid performance degradation during its instantiation. For this reasons it is recommended to have dialog flow callback on the model class itself which will avoid instantiating the class on each dialog flow evaluation.

  Note that if dialog flow is defined and it doesn't match the history the terms of the intent won't be tested at all.

  meta={'enabled': true} ^{line 3}

  Optional. Just like the most of the components in NLPCraft, the intent can have its own metadata. Intent metadata is defined as a standard JSON object which will be converted into java.util.Map instance and can be accessed in intent's terms via meta_intent() IDL function. The typical use case for declarative intent metadata is to parameterize its behavior, i.e. the behavior of its terms, with a clearly defined properties that are provided inside intent definition itself.

  term(a)={month >= 6 && !# != "z" && meta_intent('enabled') == true}[1,3] ^{line 4}
  term(b)~{ ^{line 5}
  @usrTypes = meta_model('user_types')
(# == 'order' || # == 'order_cancel') && has_all(@usrTypes, list(1, 2, 3))
}
term(a)=/org.mypackage.MyClass#termMethod/? ^{line 21}

  Term is a building block of the intent. Intent must have at least one term. Term has optional ID, a token predicate and optional quantifiers. It supports conversation context if it uses '~' symbol or not if it uses '=' symbol in its definition. For the conversational term the system will search for a match using tokens from the current request as well as the tokens from conversation STM (short-term-memory). For a non-conversational term - only tokens from the current request will be considered.

  A term is matched if its token predicate returns true. The matched term represents one or more tokens, sequential or not, that were detected in the user input. Intent has a list of terms (always at least one) that all have to be matched in the user input for the intent to match. Note that term can be optional if its min quantifier is zero. Whether the order of the terms is important for matching is governed by intent's ordered parameter.

  Term ID (a and b) is optional. It is only required by @NCIntentTerm annotation to link term's tokens to a formal parameter of the callback method. Note that term ID follows the same lexical rules as intent ID.

  Term's body can be defined in two ways:

  • IDL Expression

    Inside of curly brackets { } you can have an optional list of term variables and the mandatory term expression that must evaluate to a boolean value. Term variable name must start with @ symbol and be unique within the scope of the current term. All term variables must be defined and initialized before term expression which must be the last statement in the term:

                                        term(b)~{
                                            @a = meta_model('a')
                                            @lst = list(1, 2, 3, 4)
    
                                            has_all(@lst, list(@a, 2))
                                        }
                                    

    Term variable initialization expression as well as term's expression follow Java-like expression grammar including precedence rules, brackets and logical combinators, as well as built-in IDL functions calls:

                                        term={true} // Special case of 'constant' term.
                                        term={
                                            // Variable declarations.
                                            @a = round(1.25)
                                            @b = meta_model('my_prop')
    
                                            // Last expression must evaluate to boolean.
                                            (@a + 2) * @b > 0
                                        }
                                        term={
                                            // Variable declarations.
                                            @c = meta_tok('prop')
                                            @lst = list(1, 2, 3)
    
                                            // Last expression must evaluate to boolean.
                                            abs(@c) > 1 && size(@lst) != 5
                                        }
                                    

    NOTE: while term variable initialization expressions can have any type - the term's expression itself, i.e. the last expression in the term's body, must evaluate to a boolean result only. Failure to do so will result in a runtime exception during intent evaluation. Note also that such errors cannot be detected during intent compilation phase.

  • User-Defined Callback

    In this case the term's body is defined as a callback in a form /x.y.z.Cass#method/, where x.y.z.Class should be a fully qualified name of the class where callback is defined, and method must be the name of the callback method. This method should take one parameter of type NCTokenPredicateContext and return an instance of NCTokenPredicateResult as its result:

                                        term(a)=/org.mypackage.MyClass#termMethod/?
                                    

    Class name is optional in which case the model class will be used by default. Note that if the custom class is in fact specified, the instance of this class will be created for each term evaluation. This class must have a no-arg constructor to instantiate via standard Java reflection and its creation must be as light as possible to avoid performance degradation during its instantiation. For this reason it is recommended to have user-defined term callback on the model class itself which will avoid instantiating the class on each term evaluation.

  ? and [1,3] define an inclusive quantifier for that term, i.e. how many times the match for this term should found. You can use the following quick abbreviations:

  • * is equal to [0,∞]
  • + is equal to [1,∞]
  • ? is equal to [0,1]
  • No quantifier defaults to [1,1]

  As mentioned above the quantifier is inclusive, i.e. the [1,3] means that the term should appear once, two times or three times.

  fragment(frag, {'p1': 25, 'p2': {'a': false}}) ^{line 22}
  

  Fragment reference allows to insert the terms defined by that fragment in place of this fragment reference. Fragment reference has mandatory fragment ID parameter and optional JSON second parameter. Optional JSON parameter allows to parameterize the inserted terms' behavior and it is available to the terms via meta_frag() IDL function.

• fragment statement

  Fragments allow to group and name a set of reusable terms. Such groups can be further parameterized at the place of reference and enable the reuse of one or more terms by multiple intents. For example:

                      // Fragments.
                      fragment=buzz term~{# == meta_frag('id')}
                      fragment=when
                          term(nums)~{
                              // Term variable.
                              @type = meta_tok('nlpcraft:num:unittype')
                              @iseq = meta_tok('nlpcraft:num:isequalcondition')
  
                              # == 'nlpcraft:num' && @type == 'datetime' && @iseq == true
                          }[0,7]
  
                      // Intents.
                      intent=alarm
                          // Insert parameterized terms from fragment 'buzz'.
                          fragment(buzz, {"id": "x:alarm"})
  
                          // Insert terms from fragment 'when'.
                          fragment(when)
                  

  NOTES:

  • Fragment statements (line 2 and 3) have a name (buzz and when) and a list of terms.
  • Terms follow the same syntax as in intent definition.
  • When a fragment is referenced in intent (lines 15 and 18) it is replaced with its terms.

• import statement

  Import statement allows to import IDL declarations from either local file, classpath resource or URL:

                      // Import using absolute path.
                      import('/opt/globals.idl')
  
                      // Import using classpath resource.
                      import('org/apache/nlpcraft/examples/alarm/intents.idl')
  
                      // Import using URL.
                      import('ftp://user:password@myhost:22/opt/globals.idl')
                  

  NOTES:

  • The effect of importing is the same as if the imported declarations were inserted in place of import statement.
  • Recursive and cyclic imports are detected and safely ignored.
  • Import statement starts with import keyword and has a string parameter that indicates the location of the resource to import.
  • For the classpath resource you don't need to specify leading forward slash.

Intent Lifecycle

During NLPCraft data probe start it scans the models provided in its configuration for the intents. The scanning process goes through JSON/YAML external configurations as well as model classes when looking for IDL intents. All found intents are compiled into an internal representation before the data probe completes its start up sequence.

Note that not all intents problems can be detected at the compilation phase, and probe can start with intents not being completely validated. For example, each term in the intent must evaluate to a boolean result. This can only be checked at runtime. Another example is the number and the types of parameters passed into IDL function which is only checked at runtime as well.

Intents are compiled only once during the data probe start up sequence and cannot be re-compiled without data probe restart. Model logic, however, can affect the intent behavior through model callbacks, model metadata, user and company metadata, as well as request data all of which can change at runtime and are accessible through IDL functions.

Intent Examples

Here's few of intent examples with explanations:

Example 1:

        intent=a
            term~{# == 'x:id'}
            term(nums)~{# == 'nlpcraft:num' && lowercase(meta_tok('nlpcraft:num:unittype')) == 'datetime'}[0,2]
        

NOTES:

• Intent has ID a.
• Intent uses default conversational support (true) and default order (false).
• Intent has two conversational terms (~) that have to be found for the intent to match. Note that second term is optional as it has [0,2] quantifier.
• Both terms have to be found in the user input for the intent to match.
• First term matches any single token with ID x:id.
• Second term can appear zero, once or two times and it matches token with ID nlpcraft:num with nlpcraft:num:unittype metadata property equal to 'datetime' string.
• IDL function lowercase used on nlpcraft:num:unittype metadata property value.
• Note that since second term has ID (nums) it can be references by @NCIntentTerm annotation by the callback formal parameter.

Example 2:

        intent=id2
            flow='id1 id2'
            term={# == 'mytok' && signum(get(meta_tok('score'), 'best')) != -1}
            term={has_any(tok_groups, list('actors', 'owners')) && size(meta_part('partAlias, 'text')) > 10}
        

NOTES:

• Intent has ID id2.
• Intent has dialog flow pattern: 'id1 id2'. It expects the sequence of intents id1 and id2 somewhere in the history of previously matched intents in the course of the current conversation.
• Intent has two non-conversational terms (=). Both terms have to be present only once (their implicit quantifiers are [1,1]).
• Both terms have to be found in the user input for the intent to match.
• First term should be a token with ID mytok and have metadata property score of type map. This map should have a value with the string key 'best'. signum of this map value should not equal -1. Note that meta_tok(), get() and signum() are all built-in IDL functions.
• Second term should be a token that belongs to either actors or owners group. It should have a part token whose with alias partAlias. That part token should have metadata property text of type string, list or map. The length of this string, list or map should be greater than 10.

Syntax Highlighting

NLPCraft IDL has relatively simple syntax and you can easily configure its syntax highlighting in most modern code editors and IDEs. Here are two examples of how to add IDL syntax highlighting:

;(function()
{
    // CommonJS
    typeof(require) != 'undefined' ? SyntaxHighlighter = require('shCore').SyntaxHighlighter : null;

    function Brush()
    {
        const keywords = 'flow fragment import intent meta options term';
        const literals = 'false null true';
        const symbols =	'[\\[\\]{}*@+?~=]+';
        const fns = 'abs asin atan atan2 avg cbrt ceil comp_addr comp_city comp_country comp_id comp_name comp_postcode comp_region comp_website concat contains cos cosh count day_of_month day_of_week day_of_year degrees distinct ends_with euler exp expm1 first floor get has has_all has_any hour hypot if index_of is_alpha is_alphanum is_alphanumspace is_alphaspace is_empty is_num is_numspace is_whitespace json keys last length list log log10 log1p lowercase max meta_company meta_conv meta_frag meta_intent meta_model meta_part meta_req meta_sys meta_tok meta_user min minute month non_empty now or_else pi pow quarter radians rand replace req_addr req_agent req_id req_normtext req_tstamp reverse rint round second signum sin sinh size sort split split_trim sqrt square starts_with stdev strip substr tan tanh to_double to_int to_string tok_aliases tok_all tok_all_for_group tok_all_for_id tok_all_for_parent tok_ancestors tok_count tok_end_idx tok_find_part tok_find_parts tok_groups tok_has_part tok_id tok_index tok_is_abstract tok_is_after_group tok_is_after_id tok_is_after_parent tok_is_before_group tok_is_before_id tok_is_before_parent tok_is_bracketed tok_is_direct tok_is_english tok_is_first tok_is_freeword tok_is_last tok_is_permutated tok_is_quoted tok_is_stopword tok_is_swear tok_is_user tok_is_wordnet tok_lemma tok_parent tok_pos tok_sparsity tok_start_idx tok_stem tok_this tok_unid tok_value trim uppercase user_admin user_email user_fname user_id user_lname user_signup_tstamp values week_of_month week_of_year year';

        this.regexList = [
            { regex: SyntaxHighlighter.regexLib.singleLineCComments, css: 'comments' },	// One line comments.
            { regex: SyntaxHighlighter.regexLib.multiLineCComments,	css: 'comments' }, // Multiline comments.
            { regex: SyntaxHighlighter.regexLib.doubleQuotedString,	css: 'string' }, // String.
            { regex: SyntaxHighlighter.regexLib.singleQuotedString, css: 'string' }, // String.
            { regex: /0x[a-f0-9]+|\d+(\.\d+)?/gi, css: 'value' }, // Numbers.
            { regex: new RegExp(this.getKeywords(keywords), 'gm'), css: 'keyword' }, // Keywords.
            { regex: new RegExp(this.getKeywords(literals), 'gm'), css: 'color1' }, // Literals.
            { regex: /<|>|<=|>=|==|!=|&&|\|\|/g, css: 'color2' }, // Operators.
            { regex: new RegExp(this.getKeywords(fns), 'gm'), css: 'functions' }, // Functions.
            { regex: new RegExp(symbols, 'gm'), css: 'color3' } // Symbols.
        ];
    }

    Brush.prototype	= new SyntaxHighlighter.Highlighter();
    Brush.aliases	= ['idl'];

    SyntaxHighlighter.brushes.Idl = Brush;

    // CommonJS.
    typeof(exports) != 'undefined' ? exports.Brush = Brush : null;
})();
                

<script src="/path/to/your/scripts/shBrushIdl.js" type="text/javascript"></script>
                

<pre class="brush: idl">
    intent=xa
        flow="^(?:login)(^:logout)*$"
        meta={'enabled': true}
        term(a)={month >= 6 && # != "z" && meta_intent('enabled') == true}[1,3]
        term(b)~{
            @usrTypes = meta_model('user_types')

            (# == 'order' || # == 'order_cancel') && has_all(@usrTypes, list(1, 2, 3))
        }
</pre>
                

IDL Functions

IDL provides over 150 built-in functions that can be used in IDL intent definitions. IDL function call takes on traditional fun_name(p1, p2, ... pk) syntax form. If function has no parameters, the brackets are optional. IDL function operates on stack - its parameters are taken from the stack and its result is put back onto stack which in turn can become a parameter for the next function call and so on. IDL functions can have zero or more parameters and always have one result value. Some IDL functions support variable number of parameters. Note that you cannot define your own functions in IDL - in such cases you need to use the term with the user-defined callback method.

                tok_id() == 'id'
                tok_id == 'id' // Remember - empty parens are optional.
                # == 'id'
            

When chaining the function calls IDL uses mathematical notation (a-la Python) rather than object-oriented one: IDL length(trim(" text ")) vs. OOP-style " text ".trim().length().

IDL functions operate with the following types:

Some IDL functions are polymorphic, i.e. they can accept arguments and return result of multiple types. Encountering unsupported types will result in a runtime error during intent matching. It is especially important to watch out for the types when adding objects to various metadata containers and using that metadata in the IDL expressions.

All IDL functions are organized into the following groups:

// Result: 'true' if the current token ID is equal to 'my_id'.
tok_id == 'my_id'
# == 'my_id'
tok_id(tok_this) == 'my_id'
#(tok_this) == 'my_id'

// Result: 'true' if the current token lemma is equal to 'work'.
tok_lemma == 'work'
tok_lemma(tok_this) == 'work'

// Result: 'true' if the current token stem is equal to 'work'.
tok_stem == 'work'
tok_stem(tok_this) == 'work'

// Result: 'true' if the current token PoS tag is equal to 'NN'.
tok_pos == 'NN'
tok_pos(tok_this) == 'NN'

// Result: token sparsity value.
tok_sparsity
tok_sparsity(tok_this)

// Result: internal token globally unique ID.
tok_unid
tok_unid(tok_this)

// Result: token abstract flag.
tok_is_abstract
tok_is_abstract(tok_this)

// Result: token bracketed flag.
tok_is_bracketed
tok_is_bracketed(tok_this)

// Result: token direct flag.
tok_is_direct
tok_is_direct(tok_this)

// Result: token permutated flag.
tok_is_permutated
tok_is_permutated(tok_this)

// Result: token English detection flag.
tok_is_english
tok_is_english(tok_this)

// Result: token freeword flag.
tok_is_freeword
tok_is_freeword(tok_this)

// Result: token quoted flag.
tok_is_quoted
tok_is_quoted(tok_this)

// Result: token stopword flag.
tok_is_stopword
tok_is_stopword(tok_this)

// Result: token swear flag.
tok_is_swear
tok_is_swear(tok_this)

// Result: wether or not this is defined by user model element vs. built-in,
tok_is_user
tok_is_user(tok_this)

// Result: whether or not this token is part of WordNet dictionary.
tok_is_wordnet
tok_is_wordnet(tok_this)

// Result: list of all ancestors.
tok_ancestors
tok_ancestors(tok_this)

// Result: list of all ancestors.
tok_parent
tok_parent(tok_this)

// Result: list of groups this token belongs to.
tok_groups
tok_groups(tok_this)

// Result: the token value if this token was detected via element's value
tok_value
tok_value(tok_this)

// Result: checks if this token is known by 'alias' alias.
has(tok_aliases, 'alias')
has(tok_aliases(tok_this), 'alias')

// Result: start character index of this token in the original text.
tok_start_idx
tok_start_idx(tok_this)

// Result: end character index of this token in the original text.
tok_end_idx
tok_end_idx(tok_this)

// Result: current token.
tok_this

// Result: part token of the current token found by 'alias' alias,
//         if any, or throws runtime exception.
tok_find_part(tok_this, 'alias')

// Result: 'true' if part token of the current token found by 'alias' alias, 'false' otherwise.
tok_has_part(tok_this, 'alias')

// Result: part token 'alias' if it exists or the current token if it does not.
@this = tok_this
@tok = if(tok_has_part(@this, 'alias'), tok_find_part(@this, 'alias'), @this)

// Result: list of part tokens, potentially empty, of the current token found by 'alias' alias.
tok_find_parts(tok_this, 'alias')

// Result: part token 'alias' if it exists or the current token if it does not.
@this = tok_this
@parts = tok_find_parts(@this, 'alias')
@tok = if(is_empty(@parts), @this, first(@parts))

// Result: token original input text.
tok_txt

// Result: token normalized input text.
tok_norm_txt

// Result: 'true' if index of this token in the original input is equal to 1.
tok_index == 1
tok_index(tok_this) == 1

// Result: 'true' if this token is the first token in the original input.
tok_is_first
tok_is_first(tok_this)

// Result: 'true' if this token is the last token in the original input.
tok_is_last
tok_is_last(tok_this)

// Result: 'true' if there is a token with ID 'a' after this token.
tok_is_before_id('a')

// Result: 'true' if there is a token with ID 'a' before this token.
tok_is_after_id('a')

// Result: 'true' if this token is located after token with ID 'before' and before the token with ID 'after'.
tok_is_between_ids('before', 'after')

// Result: 'true' if this token is located after token belonging to the group 'before' and before the token belonging to the group 'after'.
tok_is_between_groups('before', 'after')

// Result: 'true' if this token is located after token with parent ID 'before' and before the token with parent ID 'after'.
tok_is_between_parents('before', 'after')

// Result: 'true' if there is a token that belongs to the group 'grp' after this token.
tok_is_before_group('grp')

// Result: 'true' if there is a token that belongs to the group 'grp' before this token.
tok_is_after_group('grp')

// Result: 'true' if there is a token with parent ID 'owner' before this token.
tok_is_after_parent('owner')

// Result: 'true' if there is a token with parent ID 'owner' after this token.
tok_is_before_parent('owner')

// Result: list of all tokens for the original input.
tok_all

// Result: number of all tokens for the original input.
tok_count

// Result: list of tokens for the original input that have ID 'id'.
tok_all_for_id('id')

// Result: list of tokens for the original input that have parent ID 'id'.
tok_all_for_parent('id')

// Result: list of tokens for the original input that belong to th group 'grp'.
tok_all_for_group('grp')

// Result: 9
length("some text")

// Result: 3
@lst = list(1, 2, 3)
size(@lst)
count(@lst)

regex('textabc', '^text.*$') // Returns 'true'.
regex('_textabc', '^text.*$') // Returns 'false'.

// Result: "text"
trim(" text ")
strip(" text ")

// Result: "TEXT"
uppercase("text")

// Result: "text"
lowercase("TeXt")

// Result: true
is_alpha("text")

// Result: true
is_alphanum("text123")

// Result: false
is_whitespace("text123")
// Result: true
is_whitespace("   ")

// Result: true
is_num("123")

// Result: true
is_numspace("  123")

// Result: true
is_alphaspace("  text  ")

// Result: true
is_alphanumspace(" 123 text  ")

// Result: [ "a", "b", "c" ]
split("a|b|c", "|")

// Result: ["a", "b", "c"]
split_trim("a | b | c", "|")

// Result: true
starts_width("abc", "ab")

// Result: true
ends_width("abc", "bc")

// Result: true
contains("abc", "bc")

// Result: 1
index_of("abc", "bc")

// Result: "bc"
substr("abc", 1, 3)

// Result: "aBC"
replace("abc", "bc", "BC")

// Result: 1.2
to_double("1.2")
// Result: 1.0
to_double(1)

// Result: 1
to_int("1.2")
to_int(1.2)

// Result: 1
abs(-1)
// Result: 1.5
abs(-1.5)

// Result: 2.0
ceil(1.5)

// Result: 1.0
floor(1.5)

// Result: 1.0
rint(1.2)

// Result: 1
round(1.2)

// Result: 1.0
signum(1.2)

// Result: 4.0
sqrt(2.0)

// Result: -3.0
cbrt(-27.0)

acos(1.0)

asin(1.0)

atan(1.0)

tan(1.0)

sin(1.0)

cos(1.0)

tanh(1.0)

sinh(1.0)

cosh(1.0)

atan2(1.0, 1.0)

degrees(1.0)

radians(1.0)

exp(1.0)

expm1(1.0)

log(1.0)

log10(1.0)

// Result: 4
square(2)

log1p(1.0)

// Result: 3.14159265359
pi

// Result: 0.5772156649
euler

// Result: 3
max(list(1, 2, 3))

// Result: 1
min(list(1, 2, 3))

// Result: 2.0
avg(list(1, 2, 3))
avg(list("1.0", 2, "3"))

stdev(list(1, 2, 3))
stdev(list("1.0", 2, "3"))

// Result: 4.0
pow(2.0, 2.0)

hypot(2.0, 2.0)

// Result: []
list
// Result: [1, 2, 3]
list(1, 2, 3)
// Result: ["1", true, 1.25]
list("1", 2 == 2, to_double('1.25'))

// Result: 1
get(list(1, 2, 3), 0)
// Result: true
get(json('{"a": true}'), "a")

// Result: true
has(list("a", "b"), "a")

// Result: false
has_all(list("a", "b"), "a")

// Result: true
has_any(list("a", "b"), list("a"))

// Result: "a"
first(list("a", "b"))

// Result: "b"
last(list("a", "b"))

// Result: ["a", "b"]
keys(json('{"a": true, "b": 1}'))

// Result: [true, 1]
values(json('{"a": true, "b": 1}'))

// Result: [3, 2, 1]
reverse(list(1, 2, 3))

// Result: [1, 2, 3]
sort(list(2, 1, 3))

// Result: false
is_empty("text")
is_empty(list(1))
is_empty(json('{"a": 1}'))

// Result: true
non_empty("text")
non_empty(list(1))
non_empty(json('{"a": 1}'))

// Result: [1, 2, 3]
distinct(list(1, 2, 2, 3, 1))

// Result: [1, 2, 3, 4]
concat(list(1, 2), list(3, 4))

// Result: 'prop' property of 'alias' part token of the current token.
meta_part('alias', 'prop')

// Result: 'nlpcraft:num:unit' token metadata property.
meta_tok('nlpcraft:num:unit')

// Result: 'my:prop' model metadata property.
meta_model('my:prop')

// Result: 'my:prop' user request data property.
meta_req('my:prop')

// Result: 'my:prop' user metadata property.
meta_user('my:prop')

// Result: 'my:prop' company metadata property.
meta_company('my:prop')

// Result: 'my:prop' intent metadata property.
meta_intent('my:prop')

// Result: 'my:prop' conversation metadata property.
meta_conv('my:prop')

// Result: 'my:prop' fragment metadata property.
meta_frag('my:prop')

// Result: 'java.home' system property.
meta_sys('java.home')
// Result: 'HOME' environment variable.
meta_sys('HOME')

// Result: 2021
year

// Result: 5
month

// Result: 5
day_of_month

// Result: 5
day_of_week

// Result: 51
day_of_year

// Result: 11
hour

// Result: 11
minute

// Result: 11
second

// Result: 2
week_of_month

// Result: 21
week_of_year

// Result: 2
quarter

// Result: 122312341212
now

// Result: server request ID.
req_id

// Result: request normalized text.
req_normtext

// Result: input receive timsstamp in ms.
req_tstamp

// Result: remote client address or 'null'.
req_addr

// Result: remote client agent or 'null'.
req_agent

// Result: user ID.
user_id

// Result: user first name.
user_fname

// Result: user last name.
user_lname

// Result: user email.
user_email

// Result: user admin flag.
user_admin

// Result: user signup timestamp in milliseconds.
user_signup_tstamp

// Result: company ID.
comp_id

// Result: company name.
comp_name

// Result: company website.
comp_website

// Result: company country.
comp_country

// Result: company region.
comp_region

// Result: company region.
comp_city

// Result: company address.
comp_addr

// Result: company postal code.
comp_postcode

// Result:
//  - 'list(1, 2, 3)' if 1st parameter is 'true'.
//  - 'null' if 1st parameter is 'false'.
if(meta_model('my_prop') == true, list(1, 2, 3), null)

// Result: Map.
json('{"a": 2, "b": [1, 2, 3]}')

// Result: "1.25"
to_string(1.25)
// Result: list("1", "2", "3")
to_string(list(1, 2, 3))

// Result: 'some_prop' model metadata or 'text' if one is 'null'.
@dflt = 'text'
or_else(meta_model('some_prop'), @dflt)

IDL Location

IDL declarations can be placed in different locations based on user preferences:

• @NCIntent annotation takes a string as its parameter that should be a valid IDL declaration. For example, Scala code snippet:

                  @NCIntent("import('/opt/myproj/global_fragments.idl')") // Importing.
                  @NCIntent("intent=act term(act)={has(tok_groups, 'act')} fragment(f1)") // Defining in place.
                  def onMatch(
                      @NCIntentTerm("act") actTok: NCToken,
                      @NCIntentTerm("loc") locToks: List[NCToken]
                  ): NCResult = {
                      ...
                  }
              

• External JSON/YAML data model configuration can provide one or more IDL declarations in intents field. For example:

                  {
                      "id": "nlpcraft.alarm.ex",
                      "name": "Alarm Example Model",
                      .
                      .
                      .
                      "intents": [
                          "import('/opt/myproj/global_fragments.idl')", // Importing.
                          "import('/opt/myproj/my_intents.idl')", // Importing.
                          "intent=alarm term~{#=='x:alarm'}" // Defining in place.
                      ]
                  }
              

• External *.idl files contain IDL declarations and can be imported in any other places where IDL declarations are allowed. See import() statement explanation below. For example:

                      /*
                       * File 'my_intents.idl'.
                       * ======================
                       */
  
                      import('/opt/globals.idl') // Import global intents and fragments.
  
                      // Fragments.
                      // ----------
                      fragment=buzz term~{# == 'x:alarm'}
                      fragment=when
                          term(nums)~{
                              // Term variables.
                              @type = meta_tok('nlpcraft:num:unittype')
                              @iseq = meta_tok('nlpcraft:num:isequalcondition')
  
                              # == 'nlpcraft:num' && @type != 'datetime' && @iseq == true
                          }[0,7]
  
                      // Intents.
                      // --------
                      intent=alarm
                          fragment(buzz)
                          fragment(when)
                  

Binding Intent

IDL intents must be bound to their callback methods. This binding is accomplished using the following Java annotations:

Here's a couple of examples of intent declarations to illustrate the basics of intent declaration and usage.

An intent from Light Switch Scala example:

            @NCIntent("intent=act term(act)={groups @@ 'act'} term(loc)={trim(id) == 'ls:loc'}*")
            @NCIntentSample(Array(
                "Turn the lights off in the entire house.",
                "Switch on the illumination in the master bedroom closet.",
                "Get the lights on.",
                "Please, put the light out in the upstairs bedroom.",
                "Set the lights on in the entire house.",
                "Turn the lights off in the guest bedroom.",
                "Could you please switch off all the lights?",
                "Dial off illumination on the 2nd floor.",
                "Please, no lights!",
                "Kill off all the lights now!",
                "No lights in the bedroom, please."
            ))
            def onMatch(
                @NCIntentTerm("act") actTok: NCToken,
                @NCIntentTerm("loc") locToks: List[NCToken]
            ): NCResult = {
                ...
            }
        

NOTES:

• The intent is defined in-place using @NCIntent annotation.
• A term match is defined as one or more tokens. Term can be optional if its min quantifier is zero.
• An intent act has two non-conversational terms: one mandatory term and another that can match zero or more tokens with method onMatch(...) as its callback.
• Terms is conversational if it uses '~' and non-conversational if it uses '=' symbol in its definition. If term is conversational, the matching algorithm will look into the conversation context short-term-memory (STM) to seek the matching tokens for this term. Note that the terms that were fully or partially matched using tokens from the conversation context will contribute a smaller weight to the overall intent matching weight since these terms are less specific. Non-conversational terms will be matched using tokens found only in the current user input without looking at the conversation context.
• Method onMatch(...) will be called if and when this intent is selected as the best match.
• Note that terms have min=1, max=1 quantifiers by default, i.e. one and only one.
• First term defines any single token that belongs to the group act. Note that model elements can belong to multiple groups.
• Second term would match zero or more tokens with ID ls:loc. Note that we use function trim on the token ID.
• Note that both terms have IDs (act and loc) that are used in onMatch(...) method parameters to automatically assign terms' tokens to the formal method parameters using @NCIntentTerm annotations.

In the following Alarm Clock Java example the intent is defined in JSON model definition and referenced in Java code using @NCIntentTerm annotation:

            {
                "id": "nlpcraft.alarm.ex",
                "name": "Alarm Example Model",
                "version": "1.0",
                "enabledBuiltInTokens": [
                    "nlpcraft:num"
                ],
                "elements": [
                    {
                        "id": "x:alarm",
                        "description": "Alarm token indicator.",
                        "synonyms": [
                            "{ping|buzz|wake|call|hit} {me|up|me up|_}",
                            "{set|_} {my|_} {wake|wake up|_} {alarm|timer|clock|buzzer|call} {up|_}"
                        ]
                    }
                ],
                "intents": [
                    "intent=alarm term~{# == 'x:alarm'} term(nums)~{# == 'nlpcraft:num' && meta_tok('nlpcraft:num:unittype') == 'datetime' && meta_tok('nlpcraft:num:isequalcondition') == true}[0,7]"
                ]
            }
        

            @NCIntentRef("alarm")
            @NCIntentSample({
                "Ping me in 3 minutes",
                "Buzz me in an hour and 15mins",
                "Set my alarm for 30s"
            })
            private NCResult onMatch(
               NCIntentMatch ctx,
               @NCIntentTerm("nums") List<NCToken> numToks
            ) {
               ...
            }
        

NOTES:

• Intent is defined in the external JSON model declaration (see line 19 in JSON file).
• This intent is referenced by annotation @NCIntentRef("alarm") with method onMatch(...) as its callback.
• This example defines an intent with two conversational terms both of which have to found for the intent to match.
• Terms is conversational if it uses '~' and non-conversational if it uses '=' symbol in its definition. If term is conversational, the matching algorithm will look into the conversation context short-term-memory (STM) to seek the matching tokens for this term. Note that the terms that were fully or partially matched using tokens from the conversation context will contribute a smaller weight to the overall intent matching weight since these terms are less specific. Non-conversational terms will be matched using tokens found only in the current user input without looking at the conversation context.
• Method onMatch(...) will be called when this intent is the best match detected.
• Note that terms have min=1, max=1 quantifiers by default.
• First term is defined as a single mandatory (min=1, max=1) user token with ID x:alarm whose element is defined in the model.
• Second term is defined as a zero or up to seven numeric built-in nlpcraft:num tokens that have unit type of datetime and are single numbers. Note that Alarm Clock model allows zero tokens in this term which would mean the current time.
• Given data model definition above the following sentences will be matched by this intent:
  • Ping me in 3 minutes
  • Buzz me in an hour and 15mins
  • Set my alarm for 30s

Intent Matching Logic

In order to understand the intent matching logic lets review the overall user request processing workflow:

• Step: 0
  

  Server receives REST call /ask or /ask/sync that contains the text of the sentence that needs to be processed.

• Step: 1
  

  At this step the server attempts to find additional variations of the input sentence by substituting certain words in the original text with synonyms from Google's BERT dataset. Note that server will not use the synonyms that are already defined in the model itself - it only tries to compensate for the potential incompleteness of the model. The result of this step is one or more sentences that all have the same meaning as the original text.

• Step: 2
  

  At this step the server takes one or more sentences from the previous step and tokenizes them. This process involves converting the text into a sequence of enriched tokens representing named entities. This step also performs the initial server-side enrichment and detection of the built-in named entities.

  The result of this step is a sequence of converted sentences, where each element is a sequence of tokens. These sequences are send down to the data probe that has requested data model deployed.

• Step: 3
  

  This is the first step of the probe-side processing. At this point the data probe receives one or more sequences of tokens. Probe then takes each sequence and performs the final enrichment by detecting user-defined elements additionally to the built-in tokens that were detected on the server during step 2 above.

• Step: 4
  

  This is an important step for understanding intent matching logic. At this step the data probe takes sequences of tokens generated at the last step and comes up with one or more parsing variants. A parsing variant is a sequence of tokens that is free from token overlapping and other parsing ambiguities. Typically, a single sequence of tokens can produce one (always) or more parsing variants.

  Let's consider the input text 'A B C D' and the following elements defined in our model:

                  "elements": [
                      {
                          "id": "elm1",
                          "synonyms": ["A B"]
                      },
                      {
                          "id": "elm2",
                          "synonyms": ["B C"]
                      },
                      {
                          "id": "elm3",
                          "synonyms": ["D"]
                      }
                  ],
                  

  All of these elements will be detected but since two of them are overlapping (elm1 and elm2) there should be two parsing variants at the output of this step:

  1. elm1('A', 'B') freeword('C') elm3('D')
  2. freeword('A') elm2('B', 'C') elm3('D')

  Note that at this point the system cannot determine which of these variants is the best one for matching - there's simply not enough information at this stage. It can only be determined when each variant is matched against model's intents - which happens in the next step.

• Step: 5
  

  At this step the actual matching between intents and variants happens. Each parsing variant from the previous step is matched against each intent. Each matching pair of a variant and an intent produce a match with a certain weight. If there are no matches at all - an error is returned. If matches were found, the match with the biggest weight is selected as a winning match. If multiple matches have the same weight, their respective variants' weights will be used to further sort them out. Finally, the intent's callback from the winning match is called.

  Although details on exact algorithm on weight calculation are too complex, here's the general guidelines on what determines the weight of the match between a parsing variant and the intent. Note that these rules coalesce around the principle idea that the more specific match always wins:

  • A match that captures more tokens has more weight than a match with less tokens. As a corollary, the match with less free words (i.e. unused words) has bigger weight than a match with more free words.
  • Tokens for user-defined elements are more important than built-in tokens.
  • A more specific match has bigger weight. In other words, a match that uses a token from the conversation context (i.e short-term-memory) has less weight than a match that only uses tokens from the current request. In the same way older tokens from the conversation give less weight than the more recent ones.

Intent Callback

Whether the intent is defined directly in @NCIntent annotation or indirectly via @NCIntentRef annotation - it is always bound to a callback method:

• Callback can only be an instance method on the class implementing NCModel interface.
• Method must have return type of NCResult.
• Method can have zero or more parameters:
  • Parameter of type NCIntentMatch, if present, must be first.
  • Any other parameters (other than the first optional NCIntentMatch) must have @NCIntentTerm annotation.
• Method must support reflection-based invocation.

@NCIntentTerm annotation marks callback parameter to receive term's tokens. This annotations can only be used for the parameters of the callbacks, i.e. methods that are annotated with @NCIntnet or @NCIntentRef. @NCIntentTerm takes a term ID as its only mandatory parameter and should be applied to callback method parameters to get the tokens associated with that term (if and when the intent was matched and that callback was invoked).

Depending on the term quantifier the method parameter type can only be one of the following types:

For example:

            @NCIntent("intent=id term(termId)~{# == 'my_token'}?")
            private NCResult onMatch(
               @NCIntentTerm("termId") Optional<NCToken> myTok
            ) {
               ...
            }
        

NOTES:

• Conversational term termId has [0,1] quantifier (it's optional).
• The formal parameter on the callback has a type of Optional<NCToken> because the term's quantifier is [0,1].
• Note that callback doesn't have an optional NCIntentMatch parameter.

NCRejection and NCIntentSkip Exceptions

There are two exceptions that can be used by intent callback logic to control intent matching process.

When NCRejection exception is thrown by the callback it indicates that user input cannot be processed as is. This exception typically indicates that user has not provided enough information in the input string to have it processed automatically. In most cases this means that the user's input is either too short or too simple, too long or too complex, missing required context, or is unrelated to the requested data model.

NCIntentSkip is a control flow exception to skip current intent. This exception can be thrown by the intent callback to indicate that current intent should be skipped (even though it was matched and its callback was called). If there's more than one intent matched the next best matching intent will be selected and its callback will be called.

This exception becomes useful when it is hard or impossible to encode the entire matching logic using only declarative IDL. In these cases the intent definition can be relaxed and the "last mile" of intent matching can happen inside of the intent callback's user logic. If it is determined that intent in fact does not match then throwing this exception allows to try next best matching intent, if any.

Note that there's a significant difference between NCIntentSkip exception and model's onMatchedIntent(...) callback. Unlike this callback, the exception does not force re-matching of all intents, it simply picks the next best intent from the list of already matched ones. The model's callback can force a full reevaluation of all intents against the user input.

NCIntentMatch Interface

NCIntentMatch interface can be passed into intent callback as its first parameter. This interface provide runtime information about the intent that was matched (i.e. the intent with which this callback was annotated with). Note also that intent context can only be the 1st parameter in the callback, and if not declared as such - it won't be passed in.

Model Callbacks

NCModel interface provides several callbacks that are invoked before, during and after intent matching. They provide an opportunity to inject user cross-cutting concerns into a standard intent matching workflow of NLPCraft. Usage of these callbacks is completely optional, yet they provide convenient joint points for logging, statistic collections, security audit and validation, explicit conversation context management, model metadata updates, and many other aspects that may depend on the standard intent matching workflow: