Visualization framework based on document representation learning

1. A visualization system, comprising:

a non-transitory memory device for storing computer readable program code; and

a processor device in communication with the memory device, the processor being operative with the computer readable program code to perform steps including

receiving a free text document,

converting the free text document into word vectors using learned word embeddings, wherein the learned word embeddings provide a measure of semantical similarity between individual words or phrases,

determining, by a trained neural network, document representations in a fixed-dimensional semantic representation space by passing the word vectors through the trained neural network, wherein more related documents lie closer than less related documents in the representation space,

applying a multi-level clustering algorithm to the document representations to generate a hierarchy of clusters, wherein at least one of the clusters is a group of document representations that tend to have likely comparisons, and

generating a visualization based on the hierarchy of clusters.

2. The system of claim 1 wherein the processor is operative with the computer readable program code to apply the multi-level clustering algorithm by applying an Infinite Mixture of Infinite Gaussian Mixtures (I2GMM) algorithm to the document representations.

3. The system of claim 1 wherein the processor is operative with the computer readable program code to generate the visualization based on the hierarchy of clusters by generating a scatter plot or tree structure of the hierarchy of clusters.

4. A computer-implemented method, comprising:

receiving a free text document;

converting the free text document into word vectors using learned word embeddings, wherein the learned word embeddings provide a measure of semantical similarity between individual words or phrases;

determining, by a trained machine learning model, document representations in a fixed-dimensional semantic representation space by passing the word vectors through the trained machine learning model, wherein more related documents lie closer than less related documents in the representation space;

applying a clustering algorithm to the document representations to generate clusters, wherein at least one of the clusters is a group of document representations that tend to have likely comparisons; and

generating a visualization based on the clusters.

5. The method of claim 4 further comprises training the machine learning model to obtain the representation space.

6. The method of claim 5 wherein training the machine learning model comprises training a neural network.

7. The method of claim 5 wherein training the machine learning model comprises training a Siamese long short-term memory (LSTM) network.

8. The method of claim 5 wherein training the machine learning model comprises:

passing pairs of documents with positive and negative labels as input to the machine learning model; and

optimizing an objective function that enforces a pair of documents with positive labels to have lower distance to each other than another pair of documents with negative labels.

9. The method of claim 5 wherein training the machine learning model comprises obtaining labels from document referral information to generate clinically relevant document similarity measures for representation learning.

10. The method of claim 4 wherein converting the free text document into the word vectors comprises passing the free text document through a trained word2vec model.

11. The method of claim 4 wherein applying the clustering algorithm to the document representations to generate the clusters comprises applying a connected component-based clustering algorithm to the document representations.

12. The method of claim 11 wherein applying the connected component-based clustering algorithm comprises:

calculating a distance matrix of the document representations;

applying a threshold to the distance matrix; and

determining connected components based on the distance matrix.

13. The method of claim 4 , wherein applying the clustering algorithm to the document representations to generate the clusters comprises applying a multi-level clustering algorithm to the document representations to generate a hierarchy of clusters.

14. The method of claim 13 where applying the multi-level clustering algorithm comprises applying an Infinite Mixture of Infinite Gaussian Mixtures (I2GMM) algorithm to the document representations.

15. The method of claim 4 wherein generating the visualization based on the clusters comprises generating a graph based on temporal order and cluster information.

16. The method of claim 4 wherein generating the visualization based on the clusters comprises generating a user interface that enables a user to select a type of document tags to be displayed in the visualization to represent documents.

17. The method of claim 4 wherein generating the visualization based on the clusters comprises projecting the document representations into a two-dimensional space represented by a scatter plot associated with a population of patients.

18. The method of claim 4 wherein generating the visualization based on the clusters comprises generating a tree structure representing documents associated with a population of patients.

19. The method of claim 18 wherein nodes of the tree structure are represented by one or more keywords extracted from associated documents.

20. One or more non-transitory computer readable media embodying a program of instructions executable by machine to perform steps comprising:

receiving a free text document;

converting the free text document into word vectors using learned word embeddings, wherein the learned word embeddings provide a measure of semantical similarity between individual words or phrases;

determining, by a trained machine learning model, document representations in a fixed-dimensional semantic representation space by passing the word vectors through the trained machine learning model, wherein more related documents lie closer than less related documents in the representation space;

applying a clustering algorithm to the document representations to generate clusters, wherein at least one of the clusters is a group of document representations that tend to have likely comparisons; and

generating a visualization based on the clusters.