Clustering data base on visual model
Some embodiments provide a non-transitory machine-readable medium that stores a program. The program receives a plurality of documents. The program further uses a visual model to generate a vector representation for each document in the plurality of documents. The program also clusters the plurality of documents into a set of clusters based on the vector representations of the plurality of documents. The program further determines a sample set of documents from the plurality of documents based on the set of clusters.
1. A non-transitory machine-readable medium storing a program executable by at least one processing unit of a device, the program comprising sets of instructions for:
receiving a plurality of documents;
generating a vector representation using a visual model for each document in the plurality of documents, wherein the visual model detects pixel values for each page of each document and propagates the pixel values through a neural network to generate the vector representation;
clustering each page of each document into a set of clusters based on the vector representations of the plurality of documents;
determining a sample set of documents from the plurality of documents based on the set of clusters, wherein the sample set of documents includes at least a first set of documents from a first set of clusters and a second set of documents from a second set of clusters;
receiving annotations for each document in the sample set of documents from an annotator; and
training an annotation AI model using the received annotations for each document in the sample set of documents.
2. The non-transitory machine-readable medium of claim 1 , wherein the program further comprises a set of instructions for converting each page of each document in the plurality of documents into an image, wherein generating the vector representation for the document comprises generating, by the visual model, a vector for each image based on the pixel values.
3. The non-transitory machine-readable medium of claim 2 , wherein clustering each page of each document into the set of clusters comprises clustering each image into the set of clusters based on the vector representations.
4. The non-transitory machine-readable medium of claim 1 , wherein the clustering comprises grouping pages in the plurality of documents having similar vector representations into a same cluster.
5. The non-transitory machine-readable medium of claim 4 , wherein a first document in the plurality of documents and a second document in the plurality of documents have similar vector representations if a cosine similarity between the vector representation of the first document and the vector representation of the second document is greater than a defined threshold value.
6. The non-transitory machine-readable medium of claim 1 , wherein determining the sample set of documents from the plurality of documents comprises randomly selecting a set of documents from each cluster in the set of clusters and including the set of documents in the sample set of documents.
7. The non-transitory machine-readable medium of claim 1 , wherein the visual model is implemented using a convolutional neural network comprising an input layer and a set of hidden layers.
8. A method comprising:
receiving a plurality of documents;
generating a vector representation using a visual model for each document in the plurality of documents, wherein the visual model detects pixel values for each page of each document and propagates the pixel values through a neural network to generate the vector representation;
clustering each page of each document into a set of clusters based on the vector representations of the plurality of documents;
determining a sample set of documents from the plurality of documents based on the set of clusters, wherein the sample set of documents includes at least a first set of documents from a first set of clusters and a second set of documents from a second set of clusters;
receiving annotations for each document in the sample set of documents from an annotator; and
training an annotation AI model using the received annotations for each document in the sample set of documents.
9. The method of claim 8 , further comprising converting each page of each document in the plurality of documents into an image, wherein generating the vector representation for the document comprises generating, by the visual model, a vector for each image based on the pixel values.
10. The method of claim 9 , wherein clustering each page of each document into the set of clusters comprises clustering each image into the set of clusters based on the vector representations.
11. The method of claim 8 , wherein the clustering comprises grouping pages in the plurality of documents having similar vector representations into a same cluster.
12. The method of claim 11 , wherein a first document in the plurality of documents and a second document in the plurality of documents have similar vector representations if a cosine similarity between the vector representation of the first document and the vector representation of the second document is greater than a defined threshold value.
13. The method of claim 8 , wherein determining the sample set of documents from the plurality of documents comprises randomly selecting a set of documents from each cluster in the set of clusters and including the set of documents in the sample set of documents.
14. The method of claim 8 , wherein the visual model is implemented using a convolutional neural network comprising an input layer and a set of hidden layers.
15. A system comprising:
a set of processing units; and
a non-transitory machine-readable medium storing instructions that when executed by at least one processing unit in the set of processing units cause the at least one processing unit to:
receive a plurality of documents;
generate a vector representation using a visual model for each document in the plurality of documents, wherein the visual model detects pixel values for each page of each document and propagates the pixel values through a neural network to generate the vector representation;
cluster each page of each document into a set of clusters based on the vector representations of the plurality of documents;
determine a sample set of documents from the plurality of documents based on the set of clusters, wherein the sample set of documents includes at least a first set of documents from a first set of clusters and a second set of documents from a second set of clusters;
receiving annotations for each document in the sample set of documents from an annotator; and
training an annotation AI model using the received annotations for each document in the sample set of documents.
16. The system of claim 15 , wherein the instructions further cause the at least one processing unit to convert each page of each document in the plurality of documents into an image, wherein generating the vector representation for the document comprises generating, by the visual model, a vector for each image based on the pixel values.
17. The system of claim 16 , wherein clustering each page of each document into the set of clusters comprises clustering each image into the set of clusters based on the vector representations.
18. The system of claim 15 , wherein the clustering comprises grouping pages in the plurality of documents having similar vector representations into a same cluster.
19. The system of claim 18 , wherein a first document in the plurality of documents and a second document in the plurality of documents have similar vector representations if a cosine similarity between the vector representation of the first document and the vector representation of the second document is greater than a defined threshold value.
20. The system of claim 15 , wherein determining the sample set of documents from the plurality of documents comprises randomly selecting a set of documents from each cluster in the set of clusters and including the set of documents in the sample set of documents.