Skip to content

Model output keys

Introduction

The previous page describes the overall structure of the output JSON object of any model block in a workflow.

Here is the description of all the keys that can be present in the output: actual keys depend on the setting of the functional parameters of the block, possibly other parameters set when the model is generated and overriding values in the input JSON.

categories

The categories array, a property of the document object, is present in the output of models performing document classification.
It is the list of categories predicted by the model block.

Each item of the array is an object representing a category, like this:

{
    "frequency": 70.62,
    "hierarchy": [
        "Sport",
        "Competition discipline",
        "Basketball"
    ],
    "id": "20000851",
    "label": "Basketball",
    "namespace": "iptc_en_1.0",
    "positions": [
        {
            "end": 14,
            "start": 0
        },
        {
            "end": 53,
            "start": 35
        },
        {
            "end": 139,
            "start": 136
        }
    ],
    "score": 4005.0,
    "winner": true

}

Based on the functional properties of the block and the version of the underlying software modules, it can have all of some of these properties:

  • id (string): category ID.
  • label (string): category label, if any.
  • hierarchy (array): in the case of a branched taxonomy, labels of all categories along the path from the root of the tree to the predicted category (included).
  • score (number): prediction score. For ML models it can also be negative. For symbolic models it ranges from 0 to infinite and is the cumulative score that was attributed to the category by the categorization rules.
  • frequency (number): relative score. For ML models is conventionally set to 100, while for symbolic models it is the percentage ratio of the category score to the sum of all categories scores. If the sum of the frequency value for all categories is less than 100 it means that the model is programmed not to return the categories with the lowest scores.
  • winner (boolean): true if the score is relatively high, false otherwise. The values of this key are affected by the ML Engine: Output winner categories only and the Only winners functional properties of the model block.

  • positions (array): the positions of the text blocks that determined the prediction. For ML models they can include the characters' ranges of sub-documents when:

    1. The Enable strict "Sub document categorization" compatibility mode option was enabled in the authoring application when generating the ML model.
    2. The ML model has been placed in the workflow in advanced mode.
    3. In the workflow the block of the ML model is preceded by the block of a symbolic model capable of identifying and returning segments.
    4. The symbolic model block has functional property Output segments enabled.
    5. The document input property of the ML model block is mapped to the document key of the symbolic model block output.
    6. The Sub-document segmentation strategy functional property of the ML model block has been set.

    The items of the positions array can also have a geometry property like this:

    {
    "end": 23419,
    "geometry": [
        {
            "box": [
                556,
                443,
                597,
                457
            ],
            "page": 7,
            "pageHeight": 1170,
            "pageWidth": 827
        }
    ],
    "score": 90,
    "start": 23402
}

    This may happen when the input to the model is the output of an Extract Converter processor block or has the same structure.

  • explanations (array): information explaining the prediction in detail. They are always present for ML models, while for symbolic models they are present only if the Output explanations functional parameter is turned on.
    Each item in the array is an object with these properties:

    • positions (array): scopes of the explanation. Each item corresponds to a scope and is an object with these properties:

      • type (string): always set to scope.
      • start (number): start position of the scope in the text.
      • end (number): end position of the scope in the text.
      • score (number): share of the prediction score due to the scope.

      • positions (array): operands of the explanation. Each item corresponds to a scope and is an object with these properties:

        • type (string): always set to operand.
        • start (number): start position of the operand in the text.
        • end (number): end position of the operand in the text.

    • ruleDetails (object): information on the symbolic rule—in the case of symbolic models—or on the algorithm to which the explanation corresponds. It is an object with these properties:

      • id: number of the rule in case of symbolic models or conventional value -1 for ML models.
      • label: details of the text feature the contributed to the prediction for ML models or the label of the rule for symbolic models, if any.

    • score (number): share of the prediction score due to the explanation.

    • type (string): name of the ML algorithm for ML models or constant rule for symbolic models.

  • namespace (string): the name of the software module carrying out the categorization.

content

The content key is a property of the document object and it's the text that has been analyzed.

document

The document key is an object that contains the results of a document analysis. It is common to all model blocks.

{
    "document": {
        analysis results
        }
}

documentData

The documentData array is a property of the document object.
It is an exact copy of the input key with the same name.

entities

The entities array is a property of the document object. It is the result of the named entity recognition activity performed by the symbolic engine.
Each item in the array represents a named entity like this:

{
    "lemma": "National Basketball Association",
    "positions": [
        {
            "end": 139,
            "start": 136
        }
    ],
    "syncon": 206693,
    "type": "ORG",
}

where:

  • The syncon and the lemma properties are respectively the outcome of the semantic analysis and lemmatization:

    • syncon is the ID of the Knowledge Graph entry corresponding to the entity.
      The value -1 means the entity was heuristically recognized since there's no Knowledge Graph entry for it.
    • lemma is the lemma—or base form—of the entity name.

  • positions is an array containing the positions of the entity occurrences in the text.

  • type is the entity type abbreviation.

extractions

The extraction array, a property of the document object, is present in the output of models performing information extraction, including thesaurus models.
It is the list of extraction predicted by the model block.

Each item of the array is an object representing an extraction, like this:

{
    "fields": [
        {
            "name": "ingredients",
            "positions": [
                {
                    "end": 502,
                    "score": 0.9969,
                    "start": 494
                }
            ],
            "score": 0.9969,
            "value": "potato"
        }
    ],
    "namespace": "hybridml",
    "template": "ingredients"
}

Based on the functional parameters of the block and the version of the underlying software modules, it can have all of some of these properties:

  • template (string): template name. In Platform extraction projects terminology the corresponding term is group, that is the group the class belongs to. It's always thesaurus for thesaurus models.

  • fields (array): each item of the array correspond to the extraction of a class and it is an object with these properties:

    • name (string): field name, that is the information class in Platform extraction projects terminology. It's always concept for thesaurus models.
    • value (string): extracted value.
    • score (number): confidence score attributed to the extraction. It is a number between 0 and 1.

    • positions (array): the positions of the text blocks that determined the extraction. Each item of the array is an object with these properties:

      • start (number): start position of the text block.
      • end (number): end position of the text block.
      • score (number): share of the prediction score due to the text block.

      • explanations (array): information explaining the prediction in detail. They are always present for ML models, while for symbolic models they are present only if the Output explanations functional parameter is turned on.
        Each item in the array is an object with these properties:

        • positions (array): rules operands or text features underlying the explanation. Each item corresponds to a text feature or rule that triggered the prediction and is an object with these properties:

          • type (string): value of the text feature for ML models.
          • start (number): start position.
          • end (number): end position.

        • ruleDetails (array): rules or types of feature. Each item is an object with these properties:

          • id: number of the rule in case of symbolic models or conventional value -1 for ML models.
          • label: the type of feature for ML models or the label of the rule for symbolic models, if any.

        • sentence (number): sentence number.

        • groupBy (number): grouping clause.
        • type (string): name of the ML algorithm for ML models or constant rule for symbolic models.

      • geometry (array): geometric information items.

  • namespace (string): name of the software module carrying out the analysis.

extraData

extraData object is s a property of the document object.

In case of a thesaurus model, it has this structure:

"extraData": {
    "thesaurusData": {}
}

If normalizeToConceptId is inserted and set to true in the API request to the workflow, then thesaurusData contains detailed information on the extracted concepts, otherwise it's empty.
The option also affects extractions: the value of extracted fields becomes a pointer to a property of the extraData object, for example:

No option or option set to false:

"extractions": [
    {
        "fields": [
            {
                name: "concept",
                value: "planet"
                ...
            }
        ...
        ],
        ...
    },
    ...
],
"extrdata": {
    "thesaurusData": {}
}

Option set to true:

"extractions": [
    {
        "fields": [
            {
                name: "concept",
                value: "12345678"
                ...
            }
        ...
        ],
        ...
    },
    ...
],
"extrdata": {
    "thesaurusData": {
        "12345678": thesaurus and project data about concept "planet",
        ...
    }
}

In case of other models, the value of extraData varies on a case-by-case basis: typically the key contains data only if the model has been produced or modified with Studio, because Studio allows producing this "extra" output via scripting.

knowledge

The knowledge array contains Knowledge Graph information about the syncons referenced, though the syncon properties of their items, in these arrays:

  • tokens
  • manSyncons
  • entities
  • relations
  • items in the sentiment object

The link between those items and the corresponding items in the knowledge array is the value of the the syncon property both have in common.
For example, if this is an item of the tokens array:

{
    "atoms": [
        {
            "end": 45,
            "lemma": "basketball",
            "start": 35,
            "type": "NOU"
        },
        {
            "end": 53,
            "lemma": "player",
            "start": 46,
            "type": "NOU"
        }
    ],
    "dependency": {
        "head": 2,
        "id": 6,
        "label": "nmod"
    },
    "end": 53,
    "lemma": "basketball player",
    "morphology": "Number=Plur",
    "paragraph": 0,
    "phrase": 2,
    "pos": "NOUN",
    "sentence": 0,
    "start": 35,
    "syncon": 41582,
    "type": "NOU"
}

the corresponding entry in the knowledge array could be:

{
    "externalIds": [
        103665646,
        43879
    ],
    "label": "person.basketball_player",
    "properties": [
        {
            "type": "DBpediaId",
            "value": "dbpedia.org/page/Basketball_player"
        },
        {
            "type": "WikiDataId",
            "value": "Q3665646"
        }
    ],
    "syncon": 41582
}

The knowledge array is a reference table: more than one item in the tokens, relations and sentiment arrays can have the same syncon ID, but there's always one entry in the knowledge array for a given syncon (it's a many-to-one relationship).
For example, if a text contains several occurrences of basketball player, each occurrence corresponds to a separate item in the tokens array, but all tokens point to the same entry in the knowledge array.

Items with the syncon property set to -1 have no corresponding entry in the knowledge array. This is because those concepts were heuristically recognized and they are not present in the Knowledge Graph, so there is no previous "knowledge" about them.

The properties of each item of the knowledge array are:

  • externalIds: array of additional identifiers of the syncon.
  • label: a textual rendering of the general conceptual category for the syncon in the Knowledge Graph.

  • The properties array: contains the references to external knowledge bases and other user data. Each item has two properties:

    • type: the name of the knowledge base or that of the user data.
    • value: the reference in thge knowledge base or the value of the user data.

    In standard Knowledge Graphs, a syncon can have reference to these knowledge bases:

    Name (value of type) Interpretation of the value
    Coordinate Latitude and longitude
    WikiDataId Wikipedia article ID
    DBpediaId URL of the DBPedia content
    GeoNamesId ID of the record in the GeoNames database

    User data are present in customized Knowledge Graphs.
    The actual list of knowledge base references and user data found in the items of the knowledge array can be set with the Required user properties for syncons functional parameter.

  • syncon: the ID of the Knowledge Graph syncon.

language

The language key, a property of the document object, is present in the output of symbolic models, symbolic steps of ML models and knowledge models.
The key value is the ISO 639-1 code of the document language.

layoutData

The layoutData key, a property of the document object, contains graphical information about the input text. This information is derived or taken from the documentLayout key of the input JSON.
layoutData is an object with these properties:

  • blocks (array): each item is an object that corresponds to a somewhat graphically distinct block of text and has these properties:

    • id (number): identification number of the block in the whole document.
    • page (number): number of the page where the block is located.

    • box (array): four items representing the coordinates1 of the bounding box of the text block:

      • Item 0: upper left corner X
      • Item 1: upper left corner Y
      • Item 2: lower right corner X
      • Item 3: lower right corner Y

    • start (number): the position of the first character of the block's text in the overall text of the document.

    • end (number): the position of the first character after the block's text in the overall text of the document.

  • fonts (array): the same information present in the homonymous key of the documentLayout input object.

  • words (array): each item is an object corresponding to a word of a text block and has these properties:

    • block (number): identification number of the block in which the word is located, it is a reference to an item of the blocks array and corresponds to the id property of that item.

    • box (array): four items representing the coordinates1 of the bounding box of the word:

      • Item 0: upper left corner X
      • Item 1: upper left corner Y
      • Item 2: lower right corner X
      • Item 3: lower right corner Y

    • start (number): the position of the first character of the word in the overall text of the document.

    • end (number): the position of the first character after the word in the overall text of the document.
    • font (number): identification number of the font with which the word is written, it is a reference to an item of the fonts array and corresponds to the id property of that item.

mainLemmas

The mainLemmas array is a property of the document object.
It contains the text main lemmas.
Each array item is an object that represents a lemma like this:

{
    "positions": [
        {
            "start": 1152,
            "end": 1162
        },
        {
            "start": 1163,
            "end": 1167
        },
        {
            "start": 1239,
            "end": 1249
        },
        {
            "start": 1335,
            "end": 1345
        },
        {
            "start": 1394,
            "end": 1404
        }
    ],
    "score": 6.5,
    "value": "locomotive"
}

where:

  • value is the lemma.
  • score is the measure of the lemma importance.
  • positions is an array containing the positions of the lemma occurrences in the text.

mainPhrases

The mainPhrases array is a property of the document object.
It contains the text main phrases.
Each array item is an object that represents a phrase like this:

{
    "positions": [
        {
            "start": 883,
            "end": 903
        }
    ],
    "score": 8,
    "value": "four-cylinder engine"
}

where:

  • value is the phrase.
  • score is the measure of the phrase importance.
  • positions is an array containing the positions of the phrase occurrences in the text.

mainSentences

The mainSentences array is a property of the document object. It contains the text main sentences.
Each array item is an object that represents a sentence like this:

{
    "end": 936,
    "score": 13.3,
    "start": 740,
    "value": "The machine is held until ready to start by a sort of trap to be sprung when all is ready; then with a tremendous flapping and snapping of the four-cylinder engine, the huge machine springs aloft."
}

where:

  • value is the sentence.
  • score is the measure of the sentence importance.
  • start is the position of the first character of the sentence.
  • end is the position of the first character after the sentence.

mainSyncons

The mainSyncons array is a property of the document object.
It contains information about the main Knowledge Graph concepts expressed in the text.
Each array item is an object that represents a Knowledge Graph concept like this:

{
    "lemma": "experiment",
    "positions": [
        {
            "end": 224,
            "start": 213
        },
        {
            "end": 2830,
            "start": 2820
        }
    ],
    "score": 5.8,
    "syncon": 2496
}

where:

  • The syncon and the lemma properties are respectively the outcome of the semantic analysis and the lemmatization.

    • syncon is the ID of the Knowledge Graph entry expressed in the text.
    • lemma is the lemma—or base form—of the concept expression (for example: scarf is the lemma for scarves).

  • score is the measure of the concept importance in the text.

  • positions is an array containing the positions of the concept occurrences in the text.

namespaces

The namespaces array is a property of the document object.
It contains resources of the software modules performing the analysis.
Each item of the array lists a namespace's resources and is an anonymous object that can have these properties:

  • categories (array): category tree. Each item of the array corresponds to a node of the category tree and it's an object with these properties:

    • id (string): category ID.
    • label (string): category label.
    • children (array): sub-tree. This property is present only if the category has subcategories. The items of the array are objects corresponding to the subcategories and have the same properties of the parent object, that is id, label and, possibly, children.

  • extractions (array): list of templates with their respective fields, corresponding to groups and classes in the terminology of Platform extraction projects. Each item is an object that corresponds to the definition of a template and has these properties:

    • name (string): template name.

    • fields (array): template fields. Each item is an object corresponding to a template's filed and has these properties:

      • name (string): field name.
      • (optional) type (string): the optional attribute of the field.

  • namespace (string): namespace identification code.

  • sections (array): list of sections. Each item of the array is an object corresponding to the definition of a section and has these properties:

    • name (string): section name.
    • score (number): numeric part of the basic section's score multiplier or 1 if the basic score is omitted in the definition of the section.

  • segments (array): list of segments. Each item of the array is an object corresponding to the definition of a segment and has this property:

    • name (string): segment name.

options

The options object is a property of the document object. It can contain the analysis, features and knowledgeProperties arrays that indicate the activities requested from NL core and the outputs it produced. Those arrays reflect output and Apply rules functional parameter of the model block or the corresponding output options in the input JSON, with some exceptions:

  • The analysis array contains the items disambiguation, entities, categories and extractions when the analysis array is missing in the input JSON. disambiguation and entities indicate that NL Core has carried out the basic analysis of the text and named entity recognition, which produce output keys content, entities, language, paragraphs, phrases, sentences, tokens and version.
    categories and extractions indicate that the categories and the extractions arrays are present in the output, but they may contain values only when the functional parameter Apply rules of the model block is turned on.
  • The analysis and features arrays are not present if the arrays of the same name in the input JSON are empty, while the knowledgeProperties array is present even if the input array of the same name is empty.

paragraphs

The paragraphs array is a property of the document object.
It contains information about the text paragraphs.
Each array item is an object that represents a paragraph like this:

{
    "end": 176,
    "sentences": [
        0,
        1
    ],
    "start": 0

}

where:

  • start is the position of the first character of the paragraph.
  • end is the position of the first character after the paragraph.
  • The sentences array contains the zero-based indexes of the constituent sentences, whose information is found in the sentences array.

phrases

The phrases array is a property of the document object.
It contains information about the text phrases.
Each array item is an object that represents a phrase like this:

{
    "end": 65,
    "start": 54,
    "tokens": [
        7,
        8,
        9
    ],
    "type": "PP"
}

where:

  • type is the phrase type. Possible phrase types are:

    Code Description
    AP Adjective Phrase
    CP Conjunction Phrase
    CR Blank lines
    DP Adverb Phrase
    NA Not Applicable (usually indicates punctuation)
    NP Noun Phrase
    PN Nominal Predicate
    PP Preposition Phrase
    RP Relative Phrase
    VP Verb Phrase

  • start is the position of the first character of the phrase.

  • end is the position of the first character after the phrase.
  • The tokens array contains the zero-based indexes of the constituent tokens, whose information is found in the tokens array.

relations

Introduction

Each item of the relations array represents a verb plus the text elements that are in a semantic relation with it. These elements may specify arguments, adjuncts or subordinate clauses. For example, given this input text:

John sent a letter to Mary.

the relations array can contain an item like this:

{
    "verb": {
        "text": "sent",
        "lemma": "send",
        "syncon": 68296,
        "phrase": 1,
        "type": "",
        "relevance": 15
    },
    "related": [
        {
            "relation": "sbj_who",
            "text": "John",
            "lemma": "John",
            "syncon": -1,
            "type": "NPH",
            "phrase": 0,
            "relevance": 15
        },
        {
            "relation": "obj_what",
            "text": "a letter",
            "lemma": "letter",
            "syncon": 29131,
            "type": "wrk",
            "phrase": 2,
            "relevance": 10
        },
        {
            "relation": "to_who",
            "text": "to Mary",
            "lemma": "Mary",
            "syncon": -1,
            "type": "NPH",
            "phrase": 3,
            "relevance": 10
        }
    ]
}

Common properties

The verb object and the items of the related array share some properties.

text is the portion of text corresponding to the element.

phrase is the index of the phrase containing the element. The value must be interpreted as a pointer to an item of the phrases array, where the positions of the first and the last character of the phrase can be found. This information can be used for text highlighting.
From the phrase, it is possible to go back to the sentence it belongs to—using the sentences array—and from the sentence to the paragraph—using the paragraphs array—or, going to the opposite direction, to find the tokens contained in the phrase —using the tokens array.
Subordinate clauses—related items having the relation property set to sub—do not have a one-to-one correspondence with a phrase. In that case, phrase has the conventional value -1.

The syncon and lemma properties are respectively the outcome of the semantic analysis and the lemmatization. Value -1 for syncon means the concept doesn't have a correspondent in the expert.ai Knowledge Graph. This can happen with:

  1. Entities having a proper noun that are heuristically recognized (for example John Smith).
  2. Parts-of-speech that are not mapped in the Knowledge Graph like pronouns (for example them).
  3. Subordinate clauses like quotes (for example John said: "I will do it!").

In cases 1 and 2, lemma is an empty string.

relevance is an indicator of the importance of the element in the text. Its value ranges from 1 to 15. When the element importance cannot be determined, relevance has the conventional value -1.

verb

The verb object is always present and it represents the verb.

type is the verb type. When set, it can be one of the following:

Verb type Description
CPL to be used as a connection as in John is a smart guy
MOV Verb of movement like to go
SAY Verb of communication like to say

The items of the related array represent text elements related to the verb.

relation is the type of relation and can be one of the following:

Possible values of relation
sbj_who
sbj_what
obj_who
obj_what
is_who
is_what
to_who
to_what
using_what
preposition* + _what
preposition* + _who
sub**
when
where
to_where
from_where
in_where
which_way
how
of_age
limited_to

* Prepositions are expressed in the language of the text intelligence engine. For example, a possible value in case of German could be auf_what. Multi-word names of prepositional expressions like according to, in front of, etc., are written in compact form without spaces between words (accordingto, infrontof).

** The sub relation type is used for subordinate clauses.

type identifies the kind of element. Possible values can be uppercase or lowercase. Uppercase corresponds to named entities, lowercase to generic entities.

Relations can be recursive: a related item can be related to another item and so on. In this case, an item of the related array can contain a related array.
For example, given this input text:

Mireille placed the plant pot on the landing at the top of the stairs.

relations can be like this:

"relations": [
    {
        "related": [
            {
                "lemma": "Mireille",
                "phrase": 0,
                "relation": "sbj_who",
                "relevance": 14,
                "syncon": -1,
                "text": "Mireille",
                "type": "NPH"
            },
            {
                "lemma": "pot",
                "phrase": 2,
                "relation": "obj_what",
                "relevance": 15,
                "syncon": 18506,
                "text": "the plant pot",
                "type": "prd"
            },
            {
                "lemma": "landing",
                "phrase": 3,
                "relation": "on_what",
                "relevance": 5,
                "syncon": 16859,
                "text": "on the landing",
                "type": "bld"
            },
            {
                "lemma": "top",
                "phrase": 4,
                "related": [
                    {
                        "lemma": "stairs",
                        "phrase": 5,
                        "relation": "of_what",
                        "relevance": 1,
                        "syncon": 20016,
                        "text": "of the stairs",
                        "type": "bld"
                    }
                ],
                "relation": "at_what",
                "relevance": -1,
                "syncon": 37732,
                "text": "at the top",
                "type": ""
            }
        ],
        "verb": {
            "lemma": "place",
            "phrase": 1,
            "relevance": 15,
            "syncon": 68498,
            "text": "placed",
            "type": ""
        }
    }
]

sections

The sections array contains the data of the text sections specified in the request, with possibly modified positions due to differences between input text and analyzed text.
Each item in the array has this format:

{
    "namespace": (string) namespace,
    "name": (string) section name,
    "positions": [
        range(s)
    ]
}

where:

  • namespace is the name of the software module carrying out document classification inside the text intelligence engine.
  • name is the name of the section.

  • The positions array indicates the range (or ranges) of characters that make up the section. Each item of the array is an object with this format:

    {
    "start": (integer) zero-based position of the first character in the section
    "end": (integer) zero-based position of the first character after the section
}

For example:

"sections": [
    {
        "namespace": "iptc_en_1.0",
        "name": "TITLE",
        "positions": [
            {
                "start": 0,
                "end": 4
            }
        ]
    },
    {
        "namespace": "iptc_en_1.0",
        "name": "BODY",
        "positions": [
            {
                "start": 6,
                "end": 10
            }
        ]
    }   
]

segments

The segments array is a property of the document object.
It contains information about the segments defined in the imported CPKs that are generated with expert.ai Studio.
It has a structure like this:

"segments": [
    {
        "name": "SEGMENT1",
        "namespace": "segments",
        "positions": [
            {
                "end": 137,
                "start": 0
            },
            {
                "end": 477,
                "start": 250
            }
        ]
    },
    {
        "name": "SEGMENT2",
        "namespace": "segments",
        "positions": [
            {
                "end": 137,
                "start": 0
            },
            {
                "end": 577,
                "start": 479
            }
        ]
    }
]

sentences

The sentences array is a property of the document object.
It contains information about the text sentences.
Each array item is an object that represents a sentence and has a structure like this:

{
    "end": 66,
    "phrases": [
        0,
        1,
        2,
        3,
        4,
        5
    ],
    "start": 0
}

where:

  • start is the position of the first character of the sentence.
  • end is the position of the first character after the sentence.
  • The phrases array contains the zero-based indexes of the constituent phrases, whose information is found in the phrases array.

sentiment

The sentiment object contains three scores indicating the tone of the whole text:

  • positivity: the amount of positivity.
  • negativity: the amount of negativity.
  • overall: the overall sentiment score, which is a combination of the scores above.

All sentiment scores are expressed in a range from -100 (extremely negative) to 100 (extremely positive).

The sentiment object contains an items array whose elements, in turn, can contain nested items arrays. These items represent the clusters of text elements that give a positive or negative contribution to the sentiment.

For example, given this input text:

The road was bad.

items clusters can be like this:

"items": [
    {
        "lemma": "road",
        "sentiment": -7,
        "syncon": 19001,
        "items": [
            {
                "lemma": "bad",
                "sentiment": -7,
                "syncon": 81195
            }
        ]
    }
]

sentiment is the sentiment score of the cluster or leaf-item. The sentiment score of a cluster is a function of the child items' scores and the possible modifiers, which are not returned as separate items, but are nevertheless taken into account.

Take, for example, a slight change introduced in the sample text:

The road was really bad.

the really modifier makes the score worse:

"items": [
    {
        "lemma": "road",
        "sentiment": -8.8,
        "syncon": 19001,
        "items": [
            {
                "lemma": "bad",
                "sentiment": -8.8,
                "syncon": 81195
            }
        ]
    }
]

On the other hand, a not before bad can invert the sentiment polarity from negative to positive. The sentiment value can be zero.

The syncon and lemma properties are respectively the outcome of the semantic analysis and the lemmatization.
An item having nested items can be an "unnamed cluster": in that case, the lemma property is an empty string.
If the intrinsic item polarity—positive or negative—is opposite to that of the paragraph it belongs to, this marker:

[*]

is added as a suffix to the the lemma.
For example, given this input text:

The road was not bad.

The lemma bad is marked with the "opposite polarity" sign because it is negated by not:

"items": [
    {
        "items": [
            {
                "lemma": "bad[*]",
                "sentiment": 7,
                "syncon": 87597
            }
        ],
        "lemma": "road",
        "sentiment": 7,
        "syncon": 19001
    }
]

Another possibility occurs when a lemma "attracts" other words in the same phrase. For example, given the input text:

Michael Jordan was one of the best basketball players of all time. Scoring was Jordan's stand-out skill, but he still holds a defensive NBA record, with eight steals in a half.

a value of lemma could be:

stand-out;skill

In this case the merged terms are separated by a semi-colon (;).

Value -1 for syncon means the concept doesn't have a correspondent in the expert.ai Knowledge Graph.

tokens

The tokens array is a property of the document object.
It contains information about the tokens in which the text was divided during the analysis.
A token is either a single word, a collocation or punctuation.
Each array item is an object that represents a token like this:

{
    "atoms": [
        {
            "end": 24,
            "lemma": "credit",
            "start": 18,
            "type": "NOU"
        },
        {
            "end": 29,
            "lemma": "card",
            "start": 25,
            "type": "NOU"
        }
    ],
    "dependency": {
        "head": 2,
        "id": 4,
        "label": "obj"
    },
    "end": 29,
    "lemma": "credit card",
    "morphology": "Number=Sing",
    "paragraph": 0,
    "phrase": 2,
    "pos": "NOUN",
    "sentence": 0,
    "start": 18,
    "syncon": 54956,
    "type": "NOU"
}

where:

  • The syncon property is the outcome of the semantic analysis process. Its value is the ID of the corresponding entry in the Knowledge Graph or -1 if there's no corresponding entry.
  • type is the type label.
  • lemma is the result of the lemmatization. It is the lemma—or base form—of the token text, for example: scarf is the lemma for scarves and be is the lemma for was.
  • pos is the result of part-of-speech tagging, the process that marks up each token with the corresponding Universal POS tag.

  • dependency is the result of syntactic analysis, the parsing process that detects the universal dependency relation between each token and the sentence root token or another token.

    The process assigns a dependency relation label to each token.
    For example, for this sentence:

    The company has developed an entirely new category of products.

    syntactic analysis determines the head token index and the dependency label as follows:

    Token index Token text Head token index Universal dependency label
    0 The 1 det
    1 company 3 nsubj
    2 has 3 aux
    3 developed 3 root
    4 an 7 det
    5 entirely 7 advmod
    6 new 7 amod
    7 category 3 obj
    8 of 9 case
    9 product 7 nmod
    10 . 3 punct

    Dependencies can be represented in various ways, such as a tree or arrow arcs.

    Inside dependency:

    • id represents the index of the token in the text.
    • dep specifies the dependency relation with another token according to the Universal Dependencies conventions.
    • head identifies the token that receives the relation. Its value corresponds to the value of the id property of another token, the only exception being the root token—the one with the depproperty set to root—for which head and id have the same value.

  • morphology is the result of morphological analysis, the process that determines lexical and grammatical features of each token in addition to the part-of-speech.

    The result of the analysis is a list of Universal features.

    For example, the morphological analysis of the first token of this sentence:

    I saw a dandelion on my lawn.

    gives:

    Case=Nom|Number=Sing|Person=1|PronType=Prs

    which is a list of feature-value pairs corresponding to:

    Pair Feature label Feature description Value label Value description
    Case=Nom Case Case Nom Nominative
    Number=Sing Number Number Sing Singular
    Person=1 Person Person 1 First
    PronType=Prs PronType Pronoun type Prs Personal

  • start is the position of the first character of the token.

  • end is the position of the first character after the token.
  • phrase is the phrase containing the token; it's the zero-based index of the phrase in the phrases array.
  • sentence is the sentence containing the token; it's the zero-based index of the sentence in the sentences array.
  • paragraph is the paragraph containing the token; it's the zero-based index of the paragraph in the paragraphs array.

  • In case of collocations—for example: credit card—, the token object can contain the atoms array. This array contains an item for every word of the collocation and has these properties:

    • type is the type label for the word.
    • lemma is the lemma of the word.
    • start is the position of the first character of the word.
    • end is the position of the first character after the word.

If the semantic analysis recognizes a token as a named entity—for example: a person's name—without a corresponding entry in the Knowledge Graph, syncon is set to -1 and the token object has an additional vsyn (virtual syncon) property like this:

{
    "syncon": -1,
    "vsyn": {
        "id": -436106,
        "parent": 73303
    },
    "start": 0,
    "end": 19,
    "type": "NPR.NPH",
    "lemma": "Mauricio Pochettino",
    ...

vsyn is an object with these properties:

  • id is a negative number assigned to all tokens considered as occurrences of the same entity. It is not the ID of a Knowledge Graph entry.
  • parent is the ID of the Knowledge Graph entry which, conceptually, is the parent of the concept expressed by the token. For example, if the token has been recognized as a person's name, parent is the ID of the concept person.

topics

The topics array is a property of the document object.
It lists the Knowledge Graph topics the text is about.
Each array item is an object that represents a Knowledge Graph topic like this:

{
    "id": 223,
    "label": "mechanics",
    "score": 3.5,
    "winner": true
}

where:

  • id is the topic ID.
  • label is the topic name.
  • score is the measure of the text topic importance.
  • winner is a boolean value set to true if the topic is considered particularly important.

version

The version key is a property of the document object.
The key value is the software module version that performed the analysis.

Type labels

The labels below are used for the type property of tokens and tokens' atoms.

Code Description
ADJ Adjective
ADV Adverb
ART Article
AUX Auxiliary verb
CON Conjunction
NOU Noun
NOU.ADR Street address
NOU.DAT Date
NOU.HOU Hour
NOU.MAI Email address
NOU.MEA Measure
NOU.MON Money
NOU.PCT Percentage
NOU.PHO Phone number
NOU.WEB Web address
NPR Proper noun
NPR.ANM Proper noun of an animal
NPR.BLD Proper noun of a building
NPR.COM Proper noun of a business/company
NPR.DEV Proper noun of a device
NPR.DOC Proper noun of a document
NPR.EVN Proper noun of an event
NPR.FDD Proper noun of a food/beverage
NPR.GEA Proper noun of a physical geographic feature
NPR.GEO Proper noun of an administrative geographic area
NPR.GEX Proper noun of an extra-terrestrial or imaginary place
NPR.LEN Proper noun of a legal/fiscal entity
NPR.MMD Proper noun of a mass media
NPR.NPH Proper noun of a human being
NPR.ORG Proper noun of an organization/society/institution
NPR.PPH Proper noun of a physical phenomena
NPR.PRD Proper noun of a product
NPR.VCL Proper noun of a vehicle
NPR.WRK Proper noun of a work of human intelligence
PNT Punctuation mark
PRE Preposition
PRO Pronoun
PRT Particle
VER Verb

Unlike Universal POS tag, used for the pos property of tokens, type labels combine part-of-speech information with entity type information and also apply to atoms.

Positions

The output of symbolic models and symbolic steps of ML models contains the position of text blocks (for example paragraphs, sentences, phrases, parts of text that "explain" predicted categories or extractions, named entities, text tokens, words, lemmas).

All these positions are referred to the content property of the document object.

The starting position is returned in the start property and the ending position in the end property.

The value of the start property is the zero-based index of the first block character.
For example, if the text is:

Michael Jordan was one of the best basketball players of all time.

the start position of the phrase of all time is 54:

Michael Jordan was one of the best basketball players of all time.
                                                      ↑
01234567890123456789012345678901234567890123456789012345678901234567890
0         1         2         3         4         5         6         7

The value of the end position is the zero-based index of the first character after the text block. In the example above, the end position of the phrase is 65:

Michael Jordan was one of the best basketball players of all time.
                                                                 ↑
01234567890123456789012345678901234567890123456789012345678901234567890
0         1         2         3         4         5         6         7

Geometry

The positions of the text that explain the predictions of the model can sometimes be enriched with geometric information.
This can occur only if the input to the model is the text of the document to be analyzed together with its graphic layout, an information that is conveyed to the model in input key documentLayout, for example when, in the workflow, the model block is preceded by a block of the Extract Converter processor.

The geometric information is like this:

"geometry": [
    {
        "box": [
            556,
            443,
            597,
            457
        ],
        "page": 7,
        "pageHeight": 1170,
        "pageWidth": 827
    }
]

where:

  • geometry is an array of geometric references, each item of which represents a rectangular area (a box) containing text and has these properties:

    • box is an array of four items representing the coordinates1 of the box:

      • Item 0: upper left corner X
      • Item 1: upper left corner Y
      • Item 2: lower right corner X
      • Item 3: lower right corner Y

    • page, pageHeight and pageWidth are, respectively, the number, height and width of the page where the box is located.

Coordinates and sizes are in pixels and referred to a 100 DPI (dots per inch) rendering of the page. The coordinates origin is at the top left corner of the rendered page.

  1. Coordinates are in pixels and referred to a 100 DPI (dots per inch) rendering of the page. The coordinates origin is at the top left corner of the rendered page. ↩↩↩