Method and system for identifying citations within regulatory content
A computer-implemented system and method for identifying citations within regulatory content is disclosed. The method involves receiving image data representing a format and layout of the regulatory content, receiving a language embedding including a plurality of tokens representing words or characters in the regulatory content, and generating a token mapping associating each of the tokens with a portion of the image data. The method also involves receiving the plurality of tokens and token mapping at an input of a citation classifier, the citation classifier having been trained to generate a classification output for each token based on the language embedding and the token mapping, the classification output identifying a plurality of citation tokens within the plurality of tokens. The method further involves processing the plurality of citation tokens to determine a hierarchical relationship between citation tokens, the hierarchical relationship being established based at least in part on the token mapping for the citation tokens.
1 . A computer-implemented method for identifying citations within regulatory content based on image data representing the regulatory content, the method comprising:
receiving a language embedding including a plurality of tokens representing words or characters in the regulatory content;
generating a token mapping associating each of the tokens with a portion of the image data representing the regulatory content;
receiving the plurality of tokens and the token mappings at an input of a citation classifier, the citation classifier having been trained to generate a classification output for each of the tokens based on the language embedding and the token mapping of that token, wherein the classification output identifies a plurality of citation tokens from the tokens; and
processing the tokens based on the classification output of each of the tokens to determine a hierarchical relationship between the tokens, the hierarchical relationship being established based at least in part on the token mapping for at least some of the tokens,
wherein processing the classification output of each of the tokens comprises processing the plurality of citation tokens to determine the hierarchical relationship between the plurality of citation tokens; and
wherein processing the plurality of citation tokens comprises receiving pairwise combinations of citation tokens at an input of a sibling classifier, the sibling classifier having been trained to output a probability indicative of whether each pairwise combination of citation tokens has a common hierarchical level.
2 . The method of claim 1 further comprising receiving text data representing characters in the regulatory content and generating the language embedding based on the received text data.
3 . The method of claim 2 wherein generating the language embedding comprises generating the language embedding using a pretrained language model to process the received text data, the pretrained language model being operably configured to generate the tokens.
4 . The method of claim 2 wherein generating the language embedding comprises:
processing the image data to identify regions of interest within the regulatory content, each region of interest including a plurality of characters; and
wherein generating the language embedding comprises generating the language embedding only for text data associated with the regions of interest.
5 . The method of claim 4 wherein processing the image data to identify the regions of interest comprises receiving the image data at an input of a region of interest classifier, the region of interest classifier having been trained to identify regions within the image data that are associated with regulatory content that should not be processed to identify citation tokens.
6 . The method of claim 5 wherein the region of interest classifier is pre-trained using a plurality of images of portions of the regulatory content that should not be processed to identify citation tokens.
7 . The method of claim 1 wherein the image data comprises a plurality of image pages and generating the token mapping comprises generating a tensor for each of the image pages, the tensor having first and second dimensions corresponding to image pixels in that image page and a third dimension in which language embedding values are associated with portions of that image page, the tensor providing the input to the citation classifier.
8 . The method of claim 1 wherein receiving the image data comprises receiving text data including format data representing a layout of the regulatory content and generating an image data representation of the text data formatted with the format data.
9 . The method of claim 1 wherein the image data comprises a plurality of page images and wherein generating the token mapping comprises:
sizing each of the page images to match a common page size;
for each of the tokens, establishing a bounding box within an applicable page image of the page images that identifies a portion of the applicable page image corresponding to that token; and
determining at least a bounding box location for each of the tokens, the bounding box location being indicative of an indentation or position of that token within the applicable page image.
10 . The method of claim 9 further comprising processing the portion of the applicable page image within the bounding box to determine at least one of a font size associated with the token or a formatting associated with the token, and wherein the citation classifier is trained to generate the classification output based on at least one of the determined font size or the determined formatting associated with the token.
11 . The method of claim 1 further comprising:
generating a similarity matrix including a plurality of rows corresponding to the plurality of citation tokens and a plurality of columns corresponding to the plurality of citation tokens, the similarity matrix being populated with the probabilities determined by the sibling classifier for each of the pairwise combinations of citation tokens; and
processing the similarity matrix to generate a hierarchical level for each of the plurality of citation tokens.
12 . The method of claim 1 wherein the plurality of citation tokens each include either a citation number or citation title, and wherein a plurality of remaining tokens not identified as the plurality of citation tokens include body text, and further comprising associating the body text between two consecutive citation tokens with a first citation token of the two consecutive citation tokens.
13 . A system for identifying citations within regulatory content based on image data representing the regulatory content, the system comprising:
a token mapper operably configured to:
receive a language embedding including a plurality of tokens representing words or characters in the regulatory content; and
generate a token mapping associating each of the tokens with a portion of the image data representing the regulatory content;
a citation classifier operably configured to receive the plurality of tokens and the token mappings at an input of a citation classifier, the citation classifier having been trained to generate a classification output for each of the tokens based on the language embedding and the token mapping of that token, wherein the classification output identifies a plurality of citation tokens within the tokens;
a hierarchical relationship classifier operably configured to process the tokens based on the classification output of each of the tokens to determine a hierarchical relationship between the tokens, the hierarchical relationship being established based at least in part on the token mapping for at least some of the tokens,
wherein the hierarchical relationship classifier is operably configured to process the tokens based on the classification output of each of the tokens by processing the plurality of citation tokens to determine the hierarchical relationship between the plurality of citation tokens; and
wherein the hierarchical relationship classifier comprises a sibling classifier and is operably configured to process the plurality of citation tokens by receiving pairwise combinations of citation tokens at an input of the sibling classifier, the sibling classifier having been trained to output a probability indicative of whether each pairwise combination of citation tokens have a common hierarchical level; and
one or more processor circuits having a memory for storing codes, the codes being operable to direct the one or more processor circuits to implement each of the token mapper, the citation classifier, and the hierarchical relationship classifier.
14 . The system of claim 13 wherein the image data comprises a plurality of page images and wherein the token mapper is further operably configured to:
size each of the page images to match a common page size;
for each of the tokens, establish a bounding box within an applicable page image of the page images that identifies a portion of the applicable page image corresponding to that token; and
determine at least a bounding box location for each of the tokens, the bounding box location being indicative of an indentation or position of that token within the applicable page image.
15 . The system of claim 13 wherein the hierarchical relationship classifier is further operably configured to:
generate a similarity matrix including a plurality of rows corresponding to the plurality of citation tokens and a plurality of columns corresponding to the plurality of citation tokens, the similarity matrix being populated with the probabilities determined by the sibling classifier for each of the pairwise combinations of citation tokens; and
process the similarity matrix to generate a hierarchical level for each of the plurality of citation tokens.