Abstract: Certain aspects of the disclosure provide techniques for web page content extraction.

Abstract: Certain aspects provide methods for training and using large language models (LLMs) to predict specialized tokens for prompting human intervention. A method includes obtaining a plurality of training data instances, each including: a training input including a first timestamp and a prompt and/or an intermediate response to the prompt, and a training output including a second timestamp and a response. The method includes annotating the training output of one or more training data instances to include an action token indicating that external intervention is required, wherein the annotation, for each training data instance, is based on: a time difference between the second timestamp and the first timestamp, a number of words included in the response; or at least one trigger word included in the response. The method also includes training the LLM on the training data instances to predict when external intervention is required and accordingly generate the action token.

Abstract: A method for automatically identifying a transactions table in a document includes obtaining a document object model of the document. The document object model includes a plurality of nodes and a plurality of edges, with each of the nodes corresponding to a respective element of the document. The method includes generating a plurality of hash values, with each of the plurality of hash values corresponding to a respective node of the plurality of nodes. The method includes determining the document includes a candidate table based on the plurality of hash values. The method includes generating a textual table based on the candidate table. The method includes analyzing one or more columns of the textual table to determine whether the textual table satisfies one or more criteria. The method includes determining the textual table is the transactions table based on determining the textual table satisfies the one or more criteria.

Abstract: Systems and methods for hardening system prompts are disclosed herein. An example method is performed by one or more processors of a hardening system. The example method may include receiving an initial prompt for a language model (LM), generating an initial accuracy score representative of an extent to which output generated by the LM matches a target output when the initial prompt is used as its system prompt, generating an initial robustness score representative of an extent to which the LM resists adversarial attacks when the initial prompt is used as its system prompt, and iteratively transforming, using an artificial intelligence (AI)-based hardening agent in conjunction with a set of machine learning (ML)-based optimization tools and a reinforcement learning (RL) technique, the initial prompt into a hardened prompt such that the hardened prompt maximizes an increase of the initial robustness score and minimizes a decrease of the initial accuracy score.

Abstract: Methods and systems are presented for providing a framework that configures a machine learning model to be insensitive to changes in input features. A computer modeling system determines data sources from which attribute values associated with transactions can be obtained. Instead of configuring the machine learning model to accept the attribute values as inputs, the computer modeling system may configure the machine learning model to accept a vector representation in a multi-dimensional space as input values. The computer modeling system then generates an encoder for each data source. Each encoder is configured to encode attribute values from a corresponding data source to a representation representing the attribute values. Further, each encoder is trained to minimize a variance between outputs of the different encoders. The computer modeling system determines a vector representation based on the representations generated by the encoders and provide the vector representation to the machine learning model.

Abstract: Methods and systems are presented for identifying different users who share a user account with an online service provider and dynamically processing transactions for the user account differently based on which user initiates the transaction request. In some embodiments, an account decomposition system may decompose the user account into distinct users who share the user account. The account decomposition system may identify different users who are sharing a user account by analyzing past transactions associated with the user account and different user devices that were used to conduct the past transactions. The account decomposition system may determine different user profiles for the different users, and may use the different user profiles to process incoming transaction requests initiated by different users of the user account.

Abstract: Certain aspects of the disclosure provide a computer-implemented method for varying machine learning model output. The method includes receiving an input for a machine learning model system; selecting, based on one or more parameters associated with the input, a set of weights of a plurality of sets of weights; providing the input to a machine learning model associated with the set of weights; and obtaining output from the machine learning model based on the input.

Abstract: A method includes classifying an uploaded document from a user application by a large language model (LLM) to obtain a ranked list of document types of the uploaded document. A multitude of pathways corresponding to the ranked list of document types is retrieved. A first portion of each of the multitude of pathways is executed in parallel to obtain a multitude of evidence structures. The evidence structures are processed using a personalized ranking model of a user of the user application, to obtain a ranked list of pathways. The ranked list of pathways is presented in the user application. A selected pathway is received from the user application. A second portion of the selected pathway is executed to process the uploaded document.

Abstract: Methods and systems are presented for identifying different users who share a user account with an online service provider and dynamically processing transactions for the user account differently based on which user initiates the transaction request. In some embodiments, an account decomposition system may decompose the user account into distinct users who share the user account. The account decomposition system may identify different users who are sharing a user account by analyzing past transactions associated with the user account and different user devices that were used to conduct the past transactions. The account decomposition system may determine different user profiles for the different users, and may use the different user profiles to process incoming transaction requests initiated by different users of the user account.

Abstract: An address encoder and a phone number encoder can be trained on training data including an address dataset, a phone number dataset, and information associating respective addresses in the address dataset with respective phone numbers in the phone number dataset as associated pairs. The training can be by a constrastive learning process such that respective distances between respective pairs of address vectors from the trained address encoder and phone number vectors from the trained phone number encoder are minimized for respective associated pairs. In production, the trained encoders can determine a production distance between a production address and a production phone number, and these results can be used to modify a production computing process.

Abstract: A method for training a machine learning model to automatically identify and extract transactions from webpages includes: obtaining sample data from a webpage, the sample data including: (i) a plurality of live transactions; and (ii) a first set of labels, each label in the first set of labels corresponding to a respective attribute of a plurality of different attributes of each of the plurality of live transactions; generating training data based on the sample data, the training data comprising: (i) a plurality of synthetic transactions; and (ii) a second set of labels including one or more labels that differ from each label included in the first set of labels; and training the machine learning model to automatically identify and extract transactions from webpages using the training data.

Abstract: Methods and systems are presented for providing a framework that configures a machine learning model to be insensitive to changes in input features. A computer modeling system determines data sources from which attribute values associated with transactions can be obtained. Instead of configuring the machine learning model to accept the attribute values as inputs, the computer modeling system may configure the machine learning model to accept a vector representation in a multi-dimensional space as input values. The computer modeling system then generates an encoder for each data source. Each encoder is configured to encode attribute values from a corresponding data source to a representation representing the attribute values. Further, each encoder is trained to minimize a variance between outputs of the different encoders. The computer modeling system determines a vector representation based on the representations generated by the encoders and provide the vector representation to the machine learning model.

Abstract: Aspects of the present disclosure relate to automated transaction categorization. Embodiments include generating, via an embedding model, a first embedding representation of a training text; assigning, via a text classification model, a class to the training text based on the first embedding representation of the training text; generating an embedding representation of a given phrase within the training text based on confirming that the class assigned to the training text is an incorrect class for the training text, wherein the given phrase is selected based on an association between the given phrase and a correct class for the training text; generating an updated embedding representation of the training text based on the first embedding representation of the training text and the embedding representation of the given phrase; and training the text classification model through a supervised learning process involving the updated embedding representation of the training text.

Abstract: A method for automatically identifying a transactions table in a document includes obtaining a document object model of the document. The document object model includes a plurality of nodes and a plurality of edges, with each of the nodes corresponding to a respective element of the document. The method includes generating a plurality of hash values, with each of the plurality of hash values corresponding to a respective node of the plurality of nodes. The method includes determining the document includes a candidate table based on the plurality of hash values. The method includes generating a textual table based on the candidate table. The method includes analyzing one or more columns of the textual table to determine whether the textual table satisfies one or more criteria. The method includes determining the textual table is the transactions table based on determining the textual table satisfies the one or more criteria.

Abstract: A method including extracting a number of page features from a web page. The number of page features represent an executable logic of the web page. The method also includes embedding, by a page feature embedding model, the number of page features to generate a page vector data structure. The method also includes comparing, by a comparison model, the page vector data structure and a number of script vector data structures to identify a selected script. Each of the number of script vector data structures is generated by a script feature embedding model processing computer executable program code of a corresponding script for performing a computer function on a web page. The method also includes presenting the selected script.

Abstract: A method including receiving a command to perform a one-to-many matching task between first and second datasets. The first and second datasets are vectorized into first and second embedded datasets. A first self-attention model is executed on the first embedded dataset to generate a first attention dataset in which each value of a first number of first features of the first dataset is weighted based on each other value of the first number of first features. A second self-attention model is executed on the second embedded dataset to generate a second attention dataset in which each value of a second number of second features of the second dataset is weighted based on each other value of the second number of second features. The first and second attention datasets are combined into a relationship matrix expressing relationships between the first and second features. The method also includes returning the relationship matrix.

Abstract: A method including applying a language model to datasets to generate topics assigned to the datasets. Each of the topics includes at least one of a natural language text word and a natural language phrase. The method also includes applying an encoding model to the topics to generate a corresponding vector data structures storing embedded topics. Each embedded topic of the embedded topics is associated with one corresponding vector in the vector data structures. The method also includes applying a clustering model to the vector data structures to generate a cluster including a subset of the vector data structures. The subset includes a reduced number of the vector data structures. The method also includes modifying, according to the cluster, the datasets.

Abstract: Systems and methods are disclosed for training a text classification model based on long text and known semantics as training data. With a text classification model limited as to the amount of text that may be input at one time, long text that is greater than the limit may be segmented into smaller segments that are less than the limit (such as into sentences). Each segment of the long text is compared with sample segments with known impact of specific semantics to associate the long text segments with the specific semantics. To compare the long text segments with sample segments, an embedding model generates an embedding from each of the segments so that the embeddings may be compared. With the long text segments associated with specific semantics, the long text segments and the associated semantics are used as training data to train a text classification model.

Abstract: Certain aspects of the disclosure provide a method for automatically identifying related communications, comprising: tokenizing a new communication from a first sender into a first plurality of tokens; identifying a plurality of communications associated with the first sender; for each respective communication: generating a self-attention data element comprising a plurality of attention values determined based on the first plurality of tokens associated with the new communication and a second plurality of tokens associated with the respective communication; determining a first plurality of features from the self-attention data element; and processing, with a machine learning (ML) model trained to identify related communications, the first plurality of features and to generate a score indicating a relatedness of the respective communication to the new communication; and determining communication(s) of the plurality of communications are related to the new communication based on each of the communication(s) hav

Abstract: Certain aspects of the disclosure provide a method for detecting data collection errors by processing error data with a plurality of regression models to generate a plurality of predicted error rates over a plurality of time intervals. The method includes determining an error mode by applying a set of policy rules optimized for determining the error mode to the plurality of predicted error rates.