Semantic map generation from natural-language text documents
Techniques include obtaining, with a computer system, a natural-language-text document comprising unstructured text; generating, with the computer system, based on a first set of machine learning model parameters, a neural representation of the unstructured text; identifying, with the computer system, based on the neural representation, a trigger word located within the unstructured text and associated with a first category; determining, with the computer system, based on the trigger word, a region within the unstructured text comprising descriptors associated with the first category; determining, with the computer system, from the region based on a second set of machine learning model parameters, a descriptor describing an action or condition of the first category; generating, with the computer system, a data model object comprising the descriptor defining an action or condition of the first category; and storing, with the computer system, the data model object in memory.
1. A tangible, non-transitory, machine-readable medium storing instructions that when executed by one or more processors effectuate operations comprising:
obtaining, with a computer system, a natural-language-text document comprising unstructured text;
generating, with the computer system, based on a first set of machine learning model parameters, a neural representation of the unstructured text, the neural representation including a sequence of a plurality of embedding vectors;
identifying, with the computer system, based on the neural representation, a trigger word located within the unstructured text and associated with a first category, wherein the trigger word is associated with the first category by using a multi-headed attention model;
determining, with the computer system, based on the trigger word, a region within the unstructured text comprising descriptors associated with the first category;
determining, with the computer system, from the region based on a second set of machine learning model parameters, a descriptor describing an action or condition of the first category;
generating, with the computer system, a data model object comprising the descriptor defining an action or condition of the first category;
storing, with the computer system, the data model object in memory; and
annotating the document with semantic triples based on the data model object, wherein the operations further comprise:
extracting, with the computer system, a document structure corresponding to the unstructured text of the natural-language-text document wherein extracting the document structure comprises:
extracting hierarchical structure within the document structure using cascading finite state machines; and
converting each branch of the document corresponding to a clause into a tagged token sequence by using a subcategorization frame template;
generating, with the computer system, based on the document structure, a directed graph representing contents of the text of the natural-language text document; and
assigning, with the computer system, based on the directed graph, a label corresponding to each word of the text to generate a tagged token sequence.
2. The medium of claim 1 , wherein the directed graph comprises:
a first graph vertex based on a first clause and associated with the data model object;
a second graph vertex based on a second clause; and
a first graph edge comprising an identifier of the data model object and an identifier of the second graph vertex.
3. The medium of claim 1 , the operations further comprising:
generating, with the computer system, an annotated document comprising annotations within the document structure, wherein vertices are annotated with descriptors corresponding to the first category; and
causing, with the computer system, a visualization of the annotated document structure in a selectable user interface (UI) to be displayed on a display device of a user.
4. The medium of claim 1 , the operations further comprising:
determining, with the computer system, based on a second trigger word and a third trigger word, a second and a third region within the unstructured text comprising descriptors associated with a second category and a third category, respectively;
determining, with the computer system, based on a second set of machine learning model parameters, a descriptor describing an action or condition of the second and third category;
wherein generating the data model object comprises:
populating a first field with the first category, wherein the first category corresponds to a right;
populating a second field with the second category, wherein the second category corresponds to an obligation; and
populating a third field with the third category, wherein the third category corresponds to a prohibition.
5. The medium of claim 1 , further comprising determining definitional statements within the text and replacing the definitional statements with lexical equivalents.
6. The medium of claim 1 , wherein each vertex of the directed graph represents a clause comprising rights, obligations, and prohibitions.
7. The medium of claim 1 , wherein each edge of the directed graph represents a relationship between clauses.
8. The medium of claim 1 , wherein the tagged token sequence is applied as an input to the first set of machine learning model parameters.
9. The medium of claim 1 , wherein generating the neural representation of the text sequence comprises replacing tokens and tags using an embedding layer to generate vector representations of each clause.
10. The medium of claim 1 , wherein the category is a right, obligation, or prohibition associated with a clause.
11. The medium of claim 10 , wherein identifying the category comprises identifying trigger words by using n-gram patterns.
12. The medium of claim 11 , the operations further comprising identifying additional categories, wherein additional categories comprise an event, a condition, a non-obligation, and a failure.
13. The medium of claim 1 , wherein determining the set of words defining an action or condition of the category comprises using a style template to compute an output score for each vector representation of clauses and selecting the vector representation having a highest output score.
14. The medium of claim 1 , wherein storing the data model object in memory comprises storing a data structure comprising the data model object in a distributed, tamper-evident data store of a peer-to-peer sharing network.
15. The medium of claim 3 , the operations further comprising:
indexing a first graph vertex based on the first category;
wherein causing the display of the visualization comprises:
retrieving the identifier of the first graph vertex based on the first category; and
displaying text stored in association with the first graph vertex.
16. The medium of claim 1 , wherein determining the data model object comprises steps for determining the data model object.
17. A method, comprising:
obtaining, with a computer system, a natural-language-text document comprising unstructured text, wherein the unstructured text comprises a plurality of clauses;
generating, with the computer system, based on a first set of machine learning model parameters, a neural representation of the unstructured text, the neural representation including a sequence of a plurality of embedding vectors;
identifying, with the computer system, a trigger word within the unstructured text and is associated with a first category, wherein the trigger word is associated with the first category by using a multi-headed attention model;
determining, with the computer system, based on the trigger word, a location within the unstructured text comprising descriptors associated with the first category;
determining, with the computer system, based on a second set of machine learning model parameters, a descriptor describing an action or condition of the first category;
generating, with the computer system, a data model object comprising the descriptor defining an action or condition of the first category;
storing, with the computer system, the data model object in memory; and
annotating the document with semantic triples based on the data model object, wherein the operations further comprise:
extracting, with the computer system, a document structure corresponding to the unstructured text of the natural-language-text document wherein extracting the document structure comprises:
extracting hierarchical structure within the document structure using cascading finite state machines; and
converting each branch of the document corresponding to a clause into a tagged token sequence by using a subcategorization frame template;
generating, with the computer system, based on the document structure, a directed graph representing contents of the text of the natural-language text document; and
assigning, with the computer system, based on the directed graph, a label corresponding to each word of the text to generate a tagged token sequence.