Cognitive framework to detect adverse events in free-form text

1. A computer-implemented method, comprising:

extracting a plurality of grammatical features of a corpus of electronic text data;

generating, based at least in part on the plurality of grammatical features, a plurality of context embedding vectors;

defining a plurality of features in a feature vector based on the plurality of context embedding vectors, comprising:

clustering the plurality of context embedding vectors, wherein a first feature of the plurality of features is defined based on a dependency between a first term and a second term in the corpus of electronic text data;

determining a first cluster identifier of a first cluster associated with the first term;

determining a second cluster identifier of a second cluster associated with the second term; and

modifying the first feature to specify a dependency between the first and second cluster identifiers;

extracting, from an input comprising unstructured electronic text, a plurality of feature values for the plurality of features defined in the feature vector;

identifying, based on a machine learning (ML) model applied to the plurality of feature values, a portion of the electronic text corresponding to an adverse event; and

annotating, by a computer processor, the identified portion of the electronic text with an indication of the adverse event.

2. The computer-implemented method of claim 1 , wherein:

extracting the plurality of grammatical features comprises processing, by a natural language processing (NLP) algorithm, the corpus of electronic text data; and

generating the plurality of context embedding vectors comprises generating, by a ML algorithm based on the plurality of grammatical features, the plurality of context embedding vectors for the corpus of electronic text data.

3. The computer-implemented method of claim 2 , wherein:

clustering the plurality of context embedding vectors comprises applying a clustering algorithm to the plurality of context embedding vectors to cluster the context embedding vectors into a plurality of clusters;

the first cluster is assigned to a first context embedding vector of the plurality of context embedding vectors, wherein the first context embedding vector is generated for the first term; and

the second cluster is assigned to a second context embedding vector of the plurality of context embedding vectors, wherein the second context embedding vector is generated for the second term.

4. The computer-implemented method of claim 3 , wherein identifying the portion of the electronic text comprising the adverse event comprises:

generating a third context embedding vector and a fourth context embedding vector for a first candidate term and a second candidate term, respectively, in the unstructured electronic text;

clustering, by the clustering algorithm, the third context embedding vector for the first candidate term into the first cluster of the plurality of clusters;

clustering, by the clustering algorithm, the fourth context embedding vector for the second candidate term into the second cluster of the plurality of clusters; and

determining, based on the clustering, that the first and second candidate terms include dependency specified by the first modified feature.

5. The computer-implemented method of claim 3 , wherein the dependency between the first term and the second term is determined based on a parse tree generated by the NLP algorithm for a first sentence of the corpus of electronic text data.

6. The computer-implemented method of claim 2 , further comprising:

training the ML model based on ML algorithm applied to the defined plurality of features and the corpus of electronic text data.

7. The computer-implemented method of claim 1 , wherein the plurality of features are further defined based on: (i) a domain specific feature, (ii) an organizational specific feature, (iii) a plurality of known adverse events specified in a dictionary.

8. A system, comprising:

a processor; and

a memory containing a program which when executed by the processor performs an operation comprising:

extracting a plurality of grammatical features of a corpus of electronic text data;

generating, based at least in part on the plurality of grammatical features, a plurality of context embedding vectors;

defining a plurality of features in a feature vector based on the plurality of context embedding vectors, comprising:

clustering the plurality of context embedding vectors, wherein a first feature of the plurality of features is defined based on a dependency between a first term and a second term in the corpus of electronic text data:

determining a first cluster identifier of a first cluster associated with the first term;

determining a second cluster identifier of a second cluster associated with the second term; and

modifying the first feature to specify a dependency between the first and second cluster identifiers;

extracting, from an input comprising unstructured electronic text, a plurality of feature values for the plurality of features defined in the feature vector;

identifying, based on a machine learning (ML) model applied to the plurality of feature values, a portion of the electronic text corresponding to an adverse event; and

annotating the identified portion of the electronic text with an indication of the adverse event.

9. The system of claim 8 , wherein:

extracting the plurality of grammatical features comprises processing, by a natural language processing (NLP) algorithm, the corpus of electronic text data; and

generating the plurality of context embedding vectors comprises generating, by a ML algorithm based on the plurality of grammatical features, the plurality of context embedding vectors for the corpus of electronic text data.

10. The system of claim 9 , wherein:

clustering the plurality of context embedding vectors comprises applying a clustering algorithm to the plurality of context embedding vectors to cluster the context embedding vectors into a plurality of clusters;

the first cluster is assigned to a first context embedding vector of the plurality of context embedding vectors, wherein the first context embedding vector is generated for the first term; and

the second cluster is assigned to a second context embedding vector of the plurality of context embedding vectors, wherein the second context embedding vector is generated for the second term.

11. The system of claim 10 , wherein identifying the portion of the electronic text comprising the adverse event comprises:

generating a third context embedding vector and a fourth context embedding vector for a first candidate term and a second candidate term, respectively, in the unstructured electronic text;

clustering, by the clustering algorithm, the third context embedding vector for the first candidate term into the first cluster of the plurality of clusters;

clustering, by the clustering algorithm, the fourth context embedding vector for the second candidate term into the second cluster of the plurality of clusters; and

determining, based on the clustering, that the first and second candidate terms include dependency specified by the first modified feature.

12. The system of claim 10 , wherein the dependency between the first term and the second term is determined based on a parse tree generated by the NLP algorithm for a first sentence of the corpus of electronic text data.

13. The system of claim 9 , the operation further comprising:

training the ML model based on ML algorithm applied to the defined plurality of features and the corpus of electronic text data.

14. The system of claim 8 , wherein the plurality of features are further defined based on: (i) a domain specific feature, (ii) an organizational specific feature, (iii) a plurality of known adverse events specified in a dictionary.

15. A computer program product, comprising:

a non-transitory computer-readable storage medium having computer-readable program code embodied therewith, the computer-readable program code executable by a processor to perform an operation comprising:

extracting a plurality of grammatical features of a corpus of electronic text data;

generating, based at least in part on the plurality of grammatical features, a plurality of context embedding vectors;

defining a plurality of features in a feature vector based on the plurality of context embedding vectors, comprising:

clustering the plurality of context embedding vectors, wherein a first feature of the plurality of features is defined based on a dependency between a first term and a second term in the corpus of electronic text data:

determining a first cluster identifier of a first cluster associated with the first term;

determining a second cluster identifier of a second cluster associated with the second term; and

modifying the first feature to specify a dependency between the first and second cluster identifiers;

extracting, from an input comprising unstructured electronic text, a plurality of feature values for the plurality of features defined in the feature vector;

identifying, based on a machine learning (ML) model applied to the plurality of feature values, a portion of the electronic text corresponding to an adverse event; and

annotating, by a computer processor, the identified portion of the electronic text with an indication of the adverse event.

16. The computer program product of claim 15 , wherein:

extracting a plurality of grammatical features comprises processing, by a natural language processing (NLP) algorithm, the corpus of electronic text data; and

generating the plurality of context embedding vectors comprises generating, by a ML algorithm based on the plurality of grammatical features, the plurality of context embedding vectors for the corpus of electronic text data.

17. The computer program product of claim 16 , wherein:

clustering the plurality of context embedding vectors comprises applying a clustering algorithm to the plurality of context embedding vectors to cluster the context embedding vectors into a plurality of clusters;

the first cluster is assigned to a first context embedding vector of the plurality of context embedding vectors, wherein the first context embedding vector is generated for the first term; and

the second cluster is assigned to a second context embedding vector of the plurality of context embedding vectors, wherein the second context embedding vector is generated for the second term.

18. The computer program product of claim 17 , wherein identifying the portion of the electronic text comprising the adverse event comprises:

generating a third context embedding vector and a fourth context embedding vector for a first candidate term and a second candidate term, respectively, in the unstructured electronic text;

clustering, by the clustering algorithm, the third context embedding vector for the first candidate term into the first cluster of the plurality of clusters;

clustering, by the clustering algorithm, the fourth context embedding vector for the second candidate term into the second cluster of the plurality of clusters; and

determining, based on the clustering, that the first and second candidate terms include dependency specified by the first modified feature.

19. The computer program product of claim 17 , wherein the dependency between the first term and the second term is determined based on a parse tree generated by the NLP algorithm for a first sentence of the corpus of electronic text data.

20. The computer program product of claim 16 , wherein the plurality of features are further defined based on: (i) a domain specific feature, (ii) an organizational specific feature, (iii) a plurality of known adverse events specified in a dictionary, the operation further comprising:

training the ML model based on ML algorithm applied to the defined plurality of features and the corpus of electronic text data.