Automatic generation of a two-part readable suspicious activity report (SAR) from high-dimensional data in tabular form
A computerized-method for automatically generating a two-part readable Suspicious Activity Report (SAR) from high-dimensional data in tabular form is provided herein. The computerized-method may include receiving high-dimensional data in tabular form of evidence financial transactions to be reported under Anti Money Laundering (AML) regulations. Then, displaying the received data to a Subject Matter Expert (SME) for ordering each displayed transaction in a predefined construction; Then, training one or more Natural Language Generation (NLG) translation models, for each transaction type, according to a deep learning model. Then, operating the one or more NLG translation models on each transaction type to generate for each transaction type a narrative of SAR; Then, operating a prebuilt summary model on the generated narrative of SAR of each transaction type to generate a summary of the plurality of narratives of SAR; and combining the plurality of narratives of SAR and the summary to one SAR.
1 . A computerized-method for automatically generating a two-part readable Suspicious Activity Report (SAR) from high-dimensional data in tabular form, said computerized-method comprising:
in a computerized-system comprising a processor and a memory, receiving by the processor, high-dimensional data in tabular form of evidence financial transactions to be reported under Anti Money Laundering (AML) regulations, performing by the processor:
displaying the received high-dimensional data in tabular form of evidence financial transactions to a Subject Matter Expert (SME) for ordering each displayed evidence financial transaction in a predefined construction;
training a plurality of Natural Language Generation (NLG) translation models on the evidence financial transactions in the predefined construction,
wherein each NLG model of the one or more NLG translation models is trained for a different preconfigured transaction type, according to a deep learning model;
selectively operating each one of the plurality of NLG translation models based on the preconfigured transaction type of the evidence financial transaction to generate a narrative of SAR for each transaction type, thus yielding a plurality of narratives of SAR, wherein each narrative of SAR is corresponding to each transaction type;
differentially weighting keywords and key sentences for each narrative of SAR in the plurality of narratives of SAR to generate a summary of the plurality of narratives of SAR by applying by the processor the deep learning model,
wherein the deep learning model differentially weight keywords and key sentences by implementing a hierarchical attention mechanism that computes attention scores of key words and key sentences in each narrative of SAR,
extracting Out Of Vocabulary (OOV) words from each narrative of SAR in the plurality of narratives of SAR for the generated summary of the plurality of narratives of SAR by applying by the processor the deep learning model,
wherein the deep learning model extracts OOV words by implementing a copying mechanism
that is based on one of: position of a word in the narrative of SAR and syntactic information of the narrative of SAR;
operating a prebuilt summary model on the differentially weighted keywords and key sentences and the extracted OOV words from each narrative of SAR in the plurality of narratives of SAR to generate the summary of the plurality of narratives of SAR; and
combining each narrative of SAR in the plurality of narratives of SAR and the generated summary to one SAR,
wherein the deep learning model is a convolutional Seq2Seq model.
2 . The computerized-method according to claim 1 , wherein the prebuilt summary model is using NLGSimple model.
3 . The computerized-method according to claim 1 , wherein each one of the different preconfigured transaction types is at least one of: international, domestic, Automated Clearing House (ACH) and Peer to Peer (P2P) transfers, and
wherein each one of the different preconfigured transaction types is performed via a channel, said channel is selected from: web, mobile, phone, branch, Application Programming Interface (API), Automated Teller Machine (ATM) and Point Of Sale (POS).
4 . The computerized-method according to claim 1 , wherein the summary is generated according to key features.
5 . The computerized-method according to claim 4 , wherein said key features are selected from: a total amount of transferred money from a first preconfigured bank to a second preconfigured bank, a total number of transfers from the first preconfigured bank to the second preconfigured bank or any other aggregated data.
6 . The computerized-method according claim 1 , wherein each evidence financial transaction is having fields categorized as high changing fields and low changing fields.
7 . The computerized-method according claim 6 , wherein high changing fields are tokenized by an attribute name, and low changing fields are tokenized by a value thereof.