Machine learning based file ranking methods and systems
A multi-label ranking method includes receiving, at a processor and from a first set of artificial neural networks (ANNs), multiple signals representing a first set of ANN output pairs for a first label. A signal representing a second set of ANN output pairs for a second label different from the first label is received at the processor from a second set of ANNs different from the first set of ANNs, substantially concurrently with the first set of ANN output pairs. A first activation function is solved based on the first set of ANN output pairs, and a second activation function is solved based on the second set of ANN output pairs. Loss values are calculated based on the solved activations, and a mask is generated based on at least one ground truth label. A signal, including a representation of the mask, is sent from the processor to each of the sets of ANNs.
1 . An apparatus, comprising:
a processor configured to execute a loss layer; and
a memory coupled to the processor, the memory storing instructions that, when executed, cause the processor to:
receive, via the loss layer and from a first plurality of artificial neural networks (ANNs), a plurality of signals representing a first plurality of ANN output pairs uniquely associated with a first label and not associated with a second label different from the first label;
receive, via the loss layer, from a second plurality of ANNs different from the first plurality of ANNs, and substantially concurrently with the first plurality of ANN output pairs, a signal representing a second plurality of ANN output pairs uniquely associated with the second label and not associated with the first label;
generate, via the loss layer, a mask by, for each portion from a plurality of portions included in the mask:
setting an indication that that portion of the mask will not cause an adjustment to a label weighting in response to detecting a lack of preference between the output pair from the first plurality of ANN output pairs associated with that portion and the output pair from the second plurality of ANN output pairs associated with that portion, and
setting an indication that that portion of the mask will cause an adjustment to the label weighting in response to detecting a preference of one of (1) the output pair from the first plurality of ANN output pairs associated with that portion and (2) the output pair from the second plurality of ANN output pairs associated with that portion; and
send a signal, including a representation of the mask, from the processor to the first plurality of ANNs and the second plurality of ANNs, such that the first plurality of ANNs and the second plurality of ANNs collectively refine a ranking model hosted by the first plurality of ANNs and the second plurality of ANNs in response to the signal.
2 . The apparatus of claim 1 , further comprising:
calculating, at the loss layer, loss values based on the first plurality of ANN output pairs and the second plurality of ANN output pairs,
wherein the adjustment to the label weighting in response to detecting the preference is associated with the loss values.
3 . The apparatus of claim 1 , further comprising:
calculating, at the loss layer, loss values using cross-entropy,
wherein the adjustment to the label weighting in response to detecting the preference is associated with the loss values.
4 . The apparatus of claim 1 , further comprising:
calculating, at the loss layer, loss values based on a ground truth associated with at least one of the first label or the second label,
wherein the adjustment to the label weighting in response to detecting the preference is associated with the loss values.
5 . The apparatus of claim 1 , wherein the adjustment to the label weighting in response to detecting the preference includes setting that portion of the mask to one.
6 . The apparatus of claim 1 , wherein the ranking model is specific to a field and configured to identify a most relevant portion of a document for the field based on user preference data.
7 . The apparatus of claim 1 , wherein at least one of the first plurality of ANNs or the second plurality of ANNs includes a feed-forward ANN, at least one of the first plurality of ANNs or the second plurality of ANNs includes a multilayer perceptron (MLP), or at least one of the first plurality of ANNs or the second plurality of ANNs includes a convolutional neural network (CNN).
8 . The apparatus of claim 1 , wherein each portion from the plurality of portions is either 0 or 1 after the generating.
9 . The apparatus of claim 1 , wherein the ranking model is configured to compare a plurality of inputs to rank relevancy of each input from the plurality of inputs to an application.
10 . A system, comprising:
a plurality of artificial neural networks (ANNs); and
a processor configured to be coupled to the plurality of ANNs and configured to execute a loss layer, the processor further configured to:
receive, via the loss layer, a plurality of ANN output pairs, each ANN output pair from the plurality of ANN output pairs associated with a different label from a plurality of labels;
generate, via the loss layer, a mask after calculating a loss value based on an activation function and based on the plurality of ANN output pairs, the generating including, for each portion from a plurality of portions included in the mask:
for each ANN output pair from the plurality of ANN output pairs and associated with that portion:
detecting whether a first ANN output of that ANN output pair is preferred over a second ANN output of that ANN output pair,
in response to detecting a lack of preference between the first ANN output and the second ANN output, setting an indication that that portion of the mask will not cause an adjustment to a label weighting for that ANN output pair, and
in response to detecting a preference of the first ANN output or the second ANN output, setting an indication that that portion of the mask will cause the adjustment to the label weighting for that ANN output pair; and
send a signal, including the mask, from the processor to a first ANN from the plurality of ANNs and a second ANN for the plurality of ANNs, such that the first ANN and the second ANN collectively update a ranking model hosted by the first ANN and the second ANN in response to the signal.
11 . The system of claim 10 , wherein the ranking model is specific to a field and configured to identify a most relevant portion of a document for the field based on user preference data.
12 . The system of claim 10 , wherein the generating the mask is based on a ground truth associated with a label from the plurality of labels.
13 . The system of claim 10 , wherein the plurality of ANN output pairs are received from a Siamese network including the first ANN and the second ANN.
14 . The system of claim 10 , wherein each label from the plurality of labels is associated with a portion of a contract.
15 . A non-transitory processor-readable medium storing instructions that, when executed, cause a processor to:
receive, via a loss layer of the processor and from a plurality of artificial neural networks (ANNs), a plurality of signals representing an associated plurality of ANN output pairs associated with a label;
define, via the loss layer, a mask, the defining including, (1) for each portion from a plurality of portions included in the mask and (2) for each ANN output pair from the plurality of ANN output pairs and associated with that portion:
detecting whether a preference exists between a first ANN output of that ANN output pair and a second ANN output of that ANN output pair,
in response to detecting a lack of preference between the first ANN output and the second ANN output, setting an indication that that portion of the mask will not cause an adjustment to a label weighting for that ANN output pair, and
in response to detecting a preference of the first ANN output or the second ANN output, setting an indication that that portion of the mask will cause the adjustment to the label weighting for that ANN output pair; and
sending a signal, including a representation of the mask, from the processor to each ANN from the plurality of ANNs, to update a ranking model of the plurality of ANNs in response to the signal.
16 . The non-transitory processor-readable medium of claim 15 , wherein the plurality of signals representing the associated plurality of ANN output pairs associated with the label is a first plurality of signals, the mask is a first mask, and the signal is a first signal, the non-transitory processor-readable medium further storing instructions that, when executed, cause a processor to:
receive, at the processor, from the plurality of ANNs, and after the first plurality of signals, a second plurality of signals representing an associated plurality of ANN output pairs associated with the label;
define, at the processor and after the first mask, a second mask based on the second plurality of signals; and
send a second signal, including the second mask, from the processor to each ANN from the plurality of ANNs, for further refinement of the ranking model of the plurality of ANNs in response to the second signal.
17 . The non-transitory processor-readable medium of claim 15 , wherein at least one ANN from the plurality of ANNs is a multilayer perceptron (MLP).
18 . The non-transitory processor-readable medium of claim 15 , wherein the ranking model is specific to a field and configured to identify a most relevant portion of a document for the field based on user preference data.
19 . The non-transitory processor-readable medium of claim 15 , wherein each ANN output pair from the plurality of ANN output pairs is generated by an associated multilayer perceptron (MLP).
20 . The non-transitory processor-readable medium of claim 15 , wherein the plurality of ANN output pairs is associated with at least two different labels and the ranking model is a single model.