Out of distribution element detection for information extraction
Techniques for extracting information from unstructured documents that enable an ML model to be trained such that the model can accurately distinguish in-distribution (“in-D”) elements and out-of-distribution (“OO-D”) elements within an unstructured document. Novel training techniques are used that train an ML model using a combination of a regular training dataset and an enhanced augmented training dataset. The regular training dataset is used to train an ML model to identify in-D elements, i.e., to classify an element extracted from a document as belonging to one of the in-D classes contained in the regular training dataset. The augmented training dataset, which is generated based upon the regular training dataset may contain one or more augmented elements which are used to train the model to identify OO-D elements, i.e., to classify an augmented element extracted from a document as belonging to an OO-D class instead of to an in-D class.
1 . A computer-implemented method comprising:
accessing a first training dataset provided for training a machine learning (ML) model, the first training dataset comprising a first plurality of documents and annotation information for each document in the first plurality of documents, wherein, for each document in the first plurality of documents, the annotation information for the document comprises information indicative of one or more elements in the document, and for each element in the one or more elements, information indicative of an in-distribution (“in-D”) class, from one or more in-D classes, to which the element belongs;
generating a second training dataset based upon the first training dataset, the second training dataset comprising a second plurality of documents and annotation information for each document in the second plurality of documents, wherein, each document in the second plurality of documents includes one or more elements that belong to an out-of-distribution (“OO-D”) class; and
training the ML model using both the first training dataset and the second training dataset to generate a trained machine learning model, wherein, for an element extracted from a particular document, the trained machine learning model is trained to classify the extracted element as belonging to an in-D class or to the OO-D class.
2 . The method of claim 1 , further comprising:
classifying, using the trained ML model, a first element extracted from the particular document as belonging to an in-D class from the one or more in-D classes; and
classifying, using the trained ML model, a second element extracted from the particular document as belonging to the OO-D class.
3 . The method of claim 1 , wherein generating the second training dataset based upon the first training dataset comprises:
identifying a first document in the first plurality of documents;
generating a second document for the second plurality of documents from the first document, wherein the second document comprises a first element; and
generating annotation information for the second document, the annotation information for the second document indicating that the first element belongs to the OO-D class.
4 . The method of claim 3 , wherein generating the second document from the first document comprises:
making a copy of the first document, wherein the second document is the copy of the first document.
5 . The method of claim 3 , wherein the first element is included in the first document.
6 . The method of claim 3 , wherein:
generating the second document from the first document comprises receiving information identifying a region within the second document;
generating the annotation information for the second document comprises including information in the annotation information for the second document indicative that any element located within the region in the second document belongs to the OO-D class; and
the first element is located in the region within the first document and also located in a region in the first document corresponding to the region in the second document.
7 . The method of claim 3 , wherein generating the second document from the first document comprises:
generating the first element;
identifying, based upon the first document, a location within the second document for placing the first element; and
placing the first element in the identified location within the second document.
8 . The method of claim 7 , wherein generating the first element comprises identifying a particular word from a set of words included in the first plurality of documents, wherein the first element is the particular word.
9 . The method of claim 7 , wherein generating the first element comprises identifying a particular word from a set of words included in the second document, wherein the first element is the particular word.
10 . The method of claim 7 , wherein generating the first element comprises:
identifying a particular word from a set of words included in the second 2 document; and
generating a similar word based upon the particular word, wherein the first element is the similar word.
11 . The method of claim 10 , wherein generating the similar word comprises:
generating an embedded representation of the particular word using a word 2 embedding model;
identifying a candidate word using a language model;
determining a similarity measure between the particular word and the candidate word; and
designating the candidate word as the similar word.
12 . The method of claim 7 , wherein generating the first element comprises:
identifying a particular word from a set of words included in the second 2 document; and
determining a synonym of the particular word, wherein the first element is the synonym.
13 . The method of claim 7 , wherein generating the first element comprises:
randomly selecting a word from a corpus of documents, wherein the first element 2 is the randomly selected word.
14 . The method of claim 7 , wherein generating the first element comprises:
identifying a particular word from a set of words included in the second document; and
identifying one or more characteristics associated with the particular word based upon metadata associated with the second document; and
generating a new word based on the one or more identified characteristics, wherein the first element is the new word.
15 . The method of claim 7 , wherein generating the first element comprises modifying a property of the first element, wherein the property includes at least one of: font, color, style, or size.
16 . The method of claim 7 , wherein generating the first element comprises generating an image, wherein the first element is the image.
17 . The method of claim 16 , wherein the image is one of a barcode, QR code, rubber stamp, handwritten text, or a watermark.
18 . A system comprising:
a set of processors;
a memory storing a trained machine learning (“ML”) model, wherein the trained ML model is trained to classify elements in a document as belonging to one of a set of one or more in-distribution (“in-D”) class or to an out-of-distribution (“OO-D”) class, wherein training the trained ML model comprises:
accessing a first training dataset, the first training dataset comprising a first plurality of documents and annotation information for each document in the first plurality of documents, wherein, for each document in the first plurality of documents, the annotation information for the document comprises information indicative of one or more elements in the document, and for each element in the one or more elements, information indicative of an in-D class to which the element belongs;
generating a second training dataset based upon the first training dataset, the second training dataset comprising a second plurality of documents and annotation information for each document in the second plurality of documents, wherein, each document in the second plurality of documents includes one or more elements that belong to the OO-D class; and
training the ML model using both the first training dataset and the second training dataset;
wherein one or more processors from the set of processors are configured to perform processing comprising:
classifying, using the trained ML model, a first element extracted from a document as belonging to an in-D class from the set of one or more in-D classes; and
classifying, using the trained ML model, a second extracted element from the document as belonging to the OO-D class.
19 . The system of claim 18 , wherein generating the second training dataset based upon the first training dataset comprises:
identifying a first document in the first plurality of documents;
generating a second document for the second plurality of documents from the first document, wherein the second document comprises a first element, comprising:
generating the first element;
identifying, based upon the first document, a location within the second document for placing the first element; and
placing the first element in the identified location within the second document; and
generating annotation information for the second document, the annotation information for the second document indicating that the first element belongs to the OO-D class.
20 . A non-transitory computer-readable medium storing computer-executable instructions that, when executed by one or more computer devices, cause the computing devices to perform processing comprising:
accessing a trained machine learning (“ML”) model, wherein the trained ML model is trained to classify elements in a document as belonging to one of a set of one or more in-distribution (“in-D”) class or to an out-of-distribution (“OO-D”) class, wherein training the trained ML model comprises:
accessing a first training dataset, the first training dataset comprising a first plurality of documents and annotation information for each document in the first plurality of documents, wherein, for each document in the first plurality of documents, the annotation information for the document comprises information indicative of one or more elements in the document, and for each element in the one or more elements, information indicative of an in-D class to which the element belongs;
generating a second training dataset based upon the first training dataset, the second training dataset comprising a second plurality of documents and annotation information for each document in the second plurality of documents, wherein, each document in the second plurality of documents includes one or more elements that belong to the OO-D class; and
training the ML model using both the first training dataset and the second training dataset;
wherein one or more processors from the set of processors are configured to perform processing comprising:
classifying, using the trained ML model, a first element extracted from a document as belonging to an in-D class from the set of one or more in-D classes; and
classifying, using the trained ML model, a second extracted element from the document as belonging to the OO-D class.