US20230176707A1 - Legal Spend Document Processing with Machine Learning - Google Patents

Abstract

A unique user interface for improving legal (and other fields) spend management is described herein. The algorithm may include looping through the lines of an invoice and displaying an icon with multiple visual indicators displaying the machine learning confidence score. When a mouse hovers over the icon, a set of icons are displayed to accept the teaching user's input. In addition, the words that drove the machine learning confidence score are highlighted with formatting so that the teaching user can understand what drove the machine learning confidence score.

Description

PRIOR APPLICATION
• This application is a continuation of U.S. patent application Ser. No. 17/744,170, “The Calibration of a Machine Learning Confidence Score”, filed on May 13, 2022, now U.S. Pat. No. 11,567,630, issued on Jan. 31, 2023, said application incorporated herein by reference in its entirety. U.S. patent application Ser. No. 17/744,170 is a continuation of U.S. patent application Ser. No. 17/328,387, “Visualization of a Machine Learning Confidence Score”, filed on May 24, 2021, now U.S. Pat. No. 11,354,018, issued on May 18, 2022, said application incorporated herein by reference in its entirety. U.S. patent application Ser. No. 17/328,387 is a continuation of U.S. patent application Ser. No. 16/919,580, “The Visualization of a Machine Learning Confidence Score and Rationale”, filed on Jul. 2, 2020, now U.S. Pat. No. 11,029,814, issued on Jun. 8, 2021, said application incorporated herein by reference in its entirety. U.S. patent application Ser. No. 16/919,580 is a continuation of U.S. patent application Ser. No. 16/299,227, “Machine Learning Visualization”, filed on Mar. 12, 2019, now U.S. Pat. No. 10,732,789, issued on Aug. 4, 2020, said application is incorporated herein by reference in its entirety. This application is also related to two design patent applications, U.S. Design patent application Ser. No. 29/678,877, “A Light Bulb Indicator of a Machine Learning Confidence Score”, filed on Jan. 31, 2019, and U.S. Design patent application Ser. No. 29/678,886, “A User Interface for Collecting Machine Learning Feedback”, filed on Jan. 31, 2019. Both applications are hereby incorporated by reference in their entirety.
BACKGROUNDTechnical Field
• The system, apparatuses and methods described herein generally relate to machine learning visualization, and, in particular, to visual techniques for displaying a machine learning confidence score and reasoning on a display screen.
Description of the Related Art
• The name machine learning was coined in 1959 by Arthur Samuel. Tom M. Mitchell provided a widely quoted, more formal definition of the algorithms studied in the machine learning field: “A computer program is said to learn from experience E with respect to some class of tasks T and performance measure P if its performance at tasks in T, as measured by P, improves with experience E.” This definition of the tasks in which machine learning is concerned offers a fundamentally operational definition rather than defining the field in cognitive terms.
• Machine learning tasks are classified into several broad categories. In supervised learning, the algorithm builds a mathematical model of a set of data that contains both the inputs and the desired outputs. For example, if the task were determining whether an image contained a certain object, the training data for a supervised learning algorithm would include images with and without that object (the input), and each image would have a label (the output) designating whether it contained the object. In special cases, the input may be only partially available or restricted to special feedback. Semi- supervised learning algorithms develop mathematical models from incomplete training data, where a portion of the sample inputs are missing the desired output.
• Machine learning implementations have two different modes, one for the operation of the algorithms and another for the learning aspect. In a supervised machine learning implementation, both the learning from training data and the actual supervised operation require that the data be specified as true or false as to the specific criteria. For instance, in the field of legal spend management, supervised machine learning is used to review legal invoices to see if the item billed fits within a set of billing criteria. If the invoice item fits within the criteria, the item is allowed. Otherwise, the invoice item is rejected. A machine learning algorithm will parse the invoice item and match certain words against a model developed by the machine learning engine. The invoice is then displayed to a user to validate the machine's determination.
• While some machine learning, Amazon storefront, for example, incorporate feedback into its user interface by taking what you buy or search for and outputting similar products as suggestions, other machine learning tasks incorporate a simple agree/disagree user interface. But neither interface provides an explanation of why the machine made its determination. There is a strong need in the industry to improve the communications from the machine to the supervising user as to why a determination is made, and to provide an easy, intuitive means for the supervising user to provide feedback to the machine.
• The present inventions address these issues.
BRIEF SUMMARY OF THE INVENTION
• A special purpose computer implemented method of visualizing a machine learning confidence score is described herein. The method is made up of the steps of processing a textual description through a machine learning model to derive the machine learning confidence score. The machine learning model uses natural language processing to convert the textual description into word stems that are used by the machine learning model to calculate the machine learning confidence score. Once the confidence score is determined, the algorithm then searches the textual description for word stems that comprised the highest impact on the confidence score. Formatting instructions are added in the textual description to at least one word associated with the word stems that comprised the highest impact on the confidence score. The textual description with the formatting instructions to indicate reasoning used by the machine learning model in the determination of the machine learning confidence score is then displayed on a display screen, and input from a user correcting the machine learning confidence score is accepted, the input used to teach the machine learning model.
• The method may use the textual description from a legal invoice, a medical invoice, or another document. A variable icon could be displayed on the display screen to graphically indicate the machine learning confidence score. The variable icon could display a different number of items depending upon the magnitude of the confidence score; in some embodiments, the variable icon displays three rays off of a lightbulb based on one range of the confidence score, and one ray off of the lightbulb based on another range of the confidence score. The variable icon could change once the input from the user is accepted. The user input could modify the machine learning model in real time.
• An apparatus for the visualization of a machine learning confidence score is also described herein. The apparatus is made up of a display screen, a special purpose computer electrically connected to the display screen, and a large capacity data storage facility with a machine learning training data set. The apparatus further includes a user interface display module operating on the special purpose computer and displaying an indication of the confidence score on the display screen; a natural language processing module operating on the special purpose computer and interfacing with the user interface display module to convert a textual description into a table of word stems; a machine learning module operating on the special purpose computer and interfacing with the natural language processing module to convert the table of word stems into a confidence score using a machine learning model built using the machine learning training data set; a scan back module operating on the special purpose computer and interfacing to the machine learning module to insert formatting instructions into the textual description using the table of stems to format at least one word in the textual description that impacted the confidence score; and a user interface input module operating on the special purpose computer and interfacing with the scan back module to display the textual description with the formatting instructions and to accept user input regarding the confidence score and to add the user input to the machine learning training data set.
• The apparatus may use the textual description from a legal invoice, a medical invoice, or another document. The indication of the confidence score on the display screen could be in the form of a variable icon that varies depending on the magnitude of the confidence score. The variable icon could display a different number of items depending upon the magnitude of the confidence score; in some embodiments, the variable icon displays three rays off of a lightbulb based on one range of the confidence score, and one ray off of the lightbulb based on another range of the confidence score. The user input could modify the machine learning model in real time.
• A system for the visualization of a machine learning confidence score is also described herein. The system is made up of a display screen on a personal computing device and a special purpose computer electrically connected to the personal computing device through a network and to a large capacity data storage facility with a machine learning training data set. The system also includes a user interface display module operating on the personal computing device and displaying the confidence score on the display screen in the form of a variable icon that varies depending on the magnitude of the confidence score; and a natural language processing module operating on the special purpose computer to convert a textual description into a table of word stems. In addition, the system includes a machine learning module operating on the special purpose computer and interfacing with the natural language processing module to convert the table of word stems into a confidence score using a machine learning model built using the machine learning training data set; a scan back module operating on the special purpose computer and interfacing to the machine learning module to insert formatting instructions into the textual description using the confidence score and the table of stems to format at least one word in the textual description that impacted the confidence score; and a user interface input module operating on the personal computing device and interfacing with the scan back module through the network to display the textual description with the formatting instructions and to accept user input regarding the confidence score and sending the user input to the special purpose computer to add the user input to the machine learning training data set.
• The system may use the textual description from a legal invoice or other documents. The variable icon could display a different number of items depending upon the magnitude of the confidence score; in some embodiments, the variable icon displays three rays off of a lightbulb based on one range of the confidence score, and one ray off of the lightbulb based on another range of the confidence score. The variable icon could change once the input from the user is accepted. The user input could modify the machine learning model in real time.
• A special purpose computer implemented method of visualizing a machine learning confidence score is recited herein. The method comprises (1) determining the machine learning confidence score on an invoice, (2) displaying an indication of the machine learning confidence score on a display screen along with the narrative section, (3) detecting a user input, (4) changing the indication of the machine learning confidence score based on the user input; and (5) displaying the word cloud. The (1) determining if the machine learning confidence score comprises (1a) parsing a narrative section of the invoice into a table of words, (1b) converting the words in the table into stems, (1c) looking up each stem in a machine learning model, (1d) placing a weight for the stem, as found in the machine learning model, in the table of the words, (1e) determining a confidence score by averaging the weights for each word in the table of the words; and looking through the narrative section and changing the formatting of the words based on a relative weight in the table of the words to create a word cloud.
• The formatting could be a bold format, color format, and/or font size. The user input could be a mouse-over and/or a touch on a touch screen. The changing of the indication of the confidence score could display a thumb icon (the thumb icon could be a thumb up icon). The method could further include (6) accepting a user indication of the acceptance of the confidence score, and possibly (7) changing the display screen to reflect that the confidence score has been accepted.
• An apparatus for visualization of a machine learning confidence score is also described herein. The apparatus comprises a display screen; a special purpose computer electrically connected to the display screen; a large capacity data storage facility with a machine learning training data set; a user interface display algorithm operates on the special purpose computer and displays an indication of the machine learning confidence score on the display screen; a natural language processing algorithm operates on the special purpose computer and interfaces with the user interface display algorithm to convert a textual description into a table of word stems; a machine learning algorithm operates on the special purpose computer and interfaces with the natural language processing algorithm to convert the table of the word stems into the machine learning confidence score using a machine learning model built using the machine learning training data set; a scan back algorithm operates on the special purpose computer and interfaces to the machine learning algorithm to insert format instructions into the textual description using the table of word stems to format at least one word in the textual description that impacted the machine learning confidence score; and a user interface input algorithm operates on the special purpose computer and interfaces with the scan back algorithm to display the textual description with the format instructions and to accept user input regarding the machine learning confidence score and to add the user input to the machine learning training data set, wherein the textual description with the format instructions is only displayed on the display screen when a mouse location is over a variable icon that varies depending on a magnitude of the machine learning confidence score.
• A special purpose computer implemented method of visualizing a machine learning confidence score is also described here. The method comprises (1) displaying an indication of the machine learning confidence score on a display screen along with a narrative section, (2) detecting that a mouse location is over the indication of the machine learning confidence score on the display screen, (3) changing the indication of the machine learning confidence score based on the mouse over the indication of the machine learning confidence score, and (4) changing the formatting of the narrative section based on the mouse over the indication of the machine learning confidence score.
• The indication of the machine learning confidence score could be a light bulb icon. The light bulb icon could vary the number of rays with the light bulb icon based on the machine learning confidence score.
• An apparatus for visualization of a machine learning confidence score is also described here. The apparatus comprises a display screen, a special purpose computer electrically connected to the display screen, a large capacity data storage facility with a machine learning training data set, a machine learning model built with the machine learning training data set, a machine learning algorithm programmed to operate on the special purpose computer and to interface with the machine learning model to convert a textual entry into the machine learning confidence score, and a user interface display algorithm programmed to operate on the special purpose computer and display an indication of the machine learning confidence score on the display screen. The indication is a variable icon that varies a count of icon elements on the variable icon depending on a magnitude of the machine learning confidence score.
• The variable icon could be a lightbulb icon, and the icon elements could be rays. The user interface display algorithm could be programmed to display three rays off of the lightbulb icon based on one range of the machine learning confidence score, and one ray off of the lightbulb icon based on a second range of the machine learning confidence score.
• The count of the icon elements could be at least three. The textual entry could be a portion of an invoice. The apparatus could also include a natural language processing algorithm programmed to operate on the special purpose computer and to interface with the user interface display algorithm to convert the textual entry into a table of word stems; and the machine learning algorithm could be programmed to operate on the special purpose computer and to interfaces with the natural language processing algorithm to convert the table of the word stems into the machine learning confidence score using the machine learning model.
• A special purpose computer implemented method of visualizing a machine learning confidence score is also described here. The method comprises the steps of (1) building a machine learning model with a machine learning training data set, (2) operating a machine learning algorithm using the machine learning model to convert a textual entry into the machine learning confidence score, (3) displaying an indication of the machine learning confidence score on a display screen, wherein the indication is a variable icon, and (4) varying a count of icon elements on the variable icon depending on a magnitude of the machine learning confidence score.
• This document also describes non-transitory machine-readable media programmed to build a machine learning model with a machine learning training data set, operate a machine learning algorithm using the machine learning model to convert a textual entry into a machine learning confidence score, display an indication of the machine learning confidence score on a display screen, wherein the indication is a variable icon, and vary a count of icon elements on the variable icon depending on a magnitude of the machine learning confidence score.
• In some aspects, the techniques described herein relate to a method of including: selecting an invoice for review by a computer, said computer electrically connected to a display screen; and looping, by the computer, through each line of the invoice: splitting the line into fields, said fields including a narrative field; processing the narrative field through a machine learning model to derive a machine learning confidence score, wherein the processing by the machine learning model includes: converting the narrative field into word stems that are used by the machine learning model to calculate the machine learning confidence score, searching the narrative field for the word stems that included a highest impact in the machine learning confidence score, and adding formatting instructions to at least one word associated with the word stems that included the highest impact on the machine learning confidence score; displaying on the display screen the fields, including the narrative field with the formatting instructions to indicate reasoning used by the machine learning model in a determination of the machine learning confidence score; and displaying a variable icon on the display screen to graphically indicate the machine learning confidence score, wherein the variable icon displays a different number of items depending upon a magnitude of the machine learning confidence score.
• In some aspects, the techniques described herein relate to a method wherein the invoice is a legal bill.
• In some aspects, the techniques described herein relate to a method further including reviewing legal spending of a lawyer as described in the invoice.
• In some aspects, the techniques described herein relate to a method wherein the reviewing includes observing the displayed narrative field and the variable icon.
• In some aspects, the techniques described herein relate to a method wherein the review is conducted by a user.
• In some aspects, the techniques described herein relate to a method wherein the invoice is a court reporter invoice.
• In some aspects, the techniques described herein relate to a method wherein the invoice is an expert witness invoice.
• In some aspects, the techniques described herein relate to a method wherein the narrative field with the formatting instructions is only displayed on the display screen when a mouse location is over an icon.
• In some aspects, the techniques described herein relate to a method wherein a bold instruction is one of the formatting instructions.
• In some aspects, the techniques described herein relate to a method wherein an instruction to increase font size is one of the formatting instructions.
• In some aspects, the techniques described herein relate to a method wherein the variable icon displays three rays off of a lightbulb based on one range of the machine learning confidence score, and one ray off of the lightbulb based on another range of the machine learning confidence score.
• In some aspects, the techniques described herein relate to a computer executing non-transitory machine readable media programmed to: select an invoice for review by the computer, said computer electrically connected to a display screen; and loop through each line of the invoice: split the line into fields, said fields including a narrative field; process the narrative field through a machine learning model to derive a machine learning confidence score, wherein the machine learning model converts the narrative field into word stems that are used by the machine learning model to calculate the machine learning confidence score, then searches the narrative field for the word stems that included a highest impact in the machine learning confidence score, and adds formatting instructions to at least one word associated with the word stems that included the highest impact on the machine learning confidence score; display on the display screen the fields, including the narrative field with the formatting instructions to indicate reasoning used by the machine learning model in a determination of the machine learning confidence score; and display a variable icon on the display screen to graphically indicate the machine learning confidence score, wherein the variable icon displays a different number of items depending upon a magnitude of the machine learning confidence score.
• In some aspects, the techniques described herein relate to a computer wherein the invoice is a legal bill.
• In some aspects, the techniques described herein relate to a computer further programmed to review legal spending of a lawyer as described in the invoice.
• In some aspects, the techniques described herein relate to a computer wherein the invoice is a court reporter invoice.
• In some aspects, the techniques described herein relate to a computer wherein the invoice is an expert witness invoice.
• In some aspects, the techniques described herein relate to a computer wherein the narrative field with the formatting instructions is only displayed on the display screen when a mouse location is over an icon.
• In some aspects, the techniques described herein relate to a computer wherein a bold instruction is one of the formatting instructions.
• In some aspects, the techniques described herein relate to a computer wherein an instruction to increase font size is one of the formatting instructions.
• In some aspects, the techniques described herein relate to a computer wherein the variable icon displays three rays off of a lightbulb based on one range of the machine learning confidence score, and one ray off of the lightbulb based on another range of the machine learning confidence score.
BRIEF DESCRIPTION OF THE DRAWINGS
• FIG.1 is a screen shot of a legal invoice with the confidence score indication.
• FIG.2 is a screen shot of a legal invoice with a mouse hovering over the confidence score indication.
• FIG.3 is a screen shot of a legal invoice with a mouse hovering over the confidence score after the user has approved the machine score.
• FIG.4 illustrates a flow chart of the user interface code in one embodiment.
• FIG.5 is a flow chart of machine learning model execution.
• FIG.6 is an electrical architecture of one embodiment.
DETAILED DESCRIPTION
• There is a strong need in the machine learning industry to provide meaningful information to the supervising user in a way that is intuitive and easy to use. Since the teaching of the machine may include thousands of evaluations from users who have little extra time to train machine models, the ease of rapidly understanding the machine's rationale and inputting feedback is critical.
• In the field of legal spend management, auditors are hired to review every legal bill sent to an insurance company or other entity to assure that the billing adheres to contractual billing guidelines. This task of reviewing invoices requires trained auditors, often lawyers themselves, performing the tedious task of reading each line of each legal bill for compliance. The reviewers vary in their interpretation due to the personalities of each auditor. In addition, the labor costs of reviewing legal bills are high.
• As a result, the legal spend management industry has started to use rule based heuristics and machine learning analysis of legal invoices to lower costs and provide a consistent interpretation. While machine learning provides an improvement over human review and heuristic determinations, in the analysis of the highly variable text in the invoices, supervised machine learning is far from easy to use.
• The present inventions provide a user interface that simply presents the machine's determination, both in terms of acceptance or rejection of an invoice item, but also provides an indication of the machine's confidence in the determination.
• While the example below uses legal invoices, the present inventions could be applied to medical bills, fleet maintenance bills, or any other type of invoices. Furthermore, any time of supervised machine learning could benefit from the user interface and algorithms described herein. Thenarrative description102 could explain the tasks of a lawyer, a court reporter, an independent adjustor, an accountant, an engineer, an expert witness, or any number of other workers.
• Looking atFIG.1, we see a portion of a legal invoice that has been processed by a machine learning model. This portion of the invoice is a table of seven rows, where each row is an individual timecard entry. Each entry has adate101, anarrative102, anamount104, and other columns. There are a number of types of analysis that can be performed that are not described herein, for instance assuring that thedates101 are within a range or that the calculations of theamount104 are correct.
• For purposes of our description in this embodiment, we focus on thenarrative102, but the other columns, or other information on the invoice or even beyond the invoice, may be used without detracting from the inventions herein. Thenarrative102 is analyzed to determine if it fits within a criterion. In the first row, the machine believes that the timecard entry in this row should be rejected, but the machine has a low confidence in its determination. This is seen because thelight bulb indicator103 has only one ray lit, indicating a low confidence. The machine's confidence score for this row is determined by the model, as described below. The machine, in its creation of the user interface, has the confidence score within a certain range that indicates a lower confidence level.
• The second row has no light bulb indicator, indicating that the machine believes that this time card entry is allowable. Technically, this means that the machine learning confidence score is below a threshold. The third row has alight bulb indicator105 with all rays lit, indicating high confidence that this time card entry should be rejected. The sixth row has alight bulb indicator107 with two rays lit, indicating medium confidence in the scoring. In one embodiment, the confidence score could have a range from 0 to 100. All confidence scores less than40 could be determined to be billable, scores from 41-60 could display oneray103, all confidence scores from 61-80 could display tworays107, and all scores above81 could display threerays105. (The number of rays could be changed without detracting from the inventions herein).
• In the fourth row, a partially disallowed timecard entry is shown, with $175.00 deducted from thisline item106.
• FIG.2 shows the display screen when the mouse hovers over thelight bulb icon205. The mouse over indicates that the user, who is teaching the machine, is focusing on the particular line of the invoice. Other embodiments could include tracking the user's eyes to determine focus or watching where a finger is hovering on a touch screen. In still another embodiment, every line of the invoice displays as in the third section ofFIG.2.
• In this example, the mouse hovers over thelight bulb icon205 and the user interface software detects the mouse over. The software then presents the user with three options to select: a thumbs upicon202, or a thumbs downicon203, or acomment icon204.
• If the user selects the thumbs upicon202, the user is approving of the machine learning decision. The time card entry in the legal invoice for that line is marked as disallowed, and the machine learning algorithm is affirmed. The thumbs upicon202 may be filled in (or a symbol (dot, for example) inserted into it) to provide feedback to the user that his input has been received. In some embodiments, a separate user interface is provided to allow the user to allow a portion of this time card entry.
• If the user selects the thumbs downicon203, the user is disapproving of the machine learning decision. The time card entry in the legal invoice for that line is marked as allowed, and the machine learning algorithm is notified that the answer is incorrect. The machine learning algorithm can adjust on the fly, or the correction can be batched up for the next run of the model. The thumbs downicon203 may be filled in (or a symbol (dot, for example) inserted into it) to provide feedback to the user that his input has been received. This is shown inFIG.3 asitem302.
• If the user selects thecomment icon204, the user is allowed to enter comments on the invoice line. Thecomment icon204 is filled in (or has a symbol (dot, for example) inserted into it) to indicate that there is a comment on this item. A second frame or window is presented to the user to collect the comment.
• Once the user has made a selection, a dot is placed in the middle of thelight bulb icon202, as seen initem304 ofFIG.3. Each time the mouse hovers over thelight bulb icon304 after the user has made a selection, the thumbs upicon202, thumbs downicon203, orcomment icon204 will indicate the previous selection by filling in the selected choice (the thumbs downicon302, for example).
• The user has the option, once a selection is made, to select theicon202,203,204 again to remove the selection. In this case, the dot in the lightbulb is removed, in one embodiment. The user may also select theother thumb icon202,203 to change the selection.
• During the mouse over, the reasoning for the machine learning confidence score is indicated in the narrative201 column for that line of theinvoice206 by changing the formatting of the words in the narrative. In the present embodiment, the words that drove the confidence score are bolded207,208 and the words that are the most important drivers of the confidence score increase theirfont sizes207. This allows the user to easily see the machine's reasoning behind the confidence score. This is similar to a teacher requiring a student to show their work as to why they made the choice that they made.
• In an alternative embodiment, the words that drive the confidence score could be displayed in a different color, italicized, single and double underlined, or any other user interface indication.
• InFIG.3, we see the display screen with the user deciding to disapprove of the machine's determination of line three of the invoice. The thumbs downicon302 is filled in and a dot is placed in the middle of thelight bulb icon304 to indicate that the user has made a selection for this item. While thethumb icons301,302 and thecomment icon303 will disappear when the mouse is no longer hovering over thelight bulb icon304, thelight bulb icon304 will continue to display the dot in the middle to indicate that the user has graded the machine's determination of this line of the invoice. In other embodiments, the dot in thelight bulb icon304 could be replaced with another type of indication, such as icon color or other visual change in theicon304.
• As inFIG.2, thenarrative line306 includes a word cloud type indication of the words that drove the machine learning confidence score. In addition, the user may note305 a partial or full dismissal of the invoice amount for that line item as well as a categorization of the reason for cutting the bill.
• FIG.4 shows one possible user interface algorithm for implementing the above described user interface. The algorithm begins when the user (or an automaton) selects an invoice to review401. The invoice is then processed by themachine learning model402. Each line of the invoice is analyzed to determine a machine learning confidence score for the line. Based on the confidence score, different icons are displayed on the screen to provide the user with information on the machine's determination403. In one embodiment, the confidence score could have a range from 0 to 100. All confidence scores less than 40 could be determined to be billable, with no icon displayed, scores from 41-60 could display oneray103, all confidence scores from 61-80 could display tworays107, and all scores above 81 could display threerays105. (The number of rays and categories could be changed without detracting from the inventions herein).
• The mouse location is then checked by the user interface404, and if the mouse is over alight bulb icon103,105,107, then the line is processed to indicate details of the reasoning. The word cloud of the narrative text is determined405 (seeFIG.5) and displayed406, replacing the text of the narrative with the narrative text with formatting indicating the reasoning for the machine learning score. The thumbs upicon202, thumbs downicon203, andcomment icon204 are also displayed where the mouse is located.
• If the mouse moves to another area of the screen, the formatting of the text narrative and thumbsicons202,203,204 will be removed from the screen.
• If the user selects one of thethumbs icons202,203,204, the screen will be updated408 to reflect the choice407. The selected icon will be filled in (thumbs downicon302, for example) and a dot will be placed in thelight bulb icon304. The user's teaching indication will then be sent back to themachine learning model409. The invoice is then revised if both the machine and the user agree that the billing entry needs to be rejected410.
• In some embodiments, when the invoice is run through themachine learning model402, a data structure is returned for each line in the invoice with the narrative formatted to indicate the importance of each word in the narrative to the determination of the confidence score. In this scenario, the functionality of405 is incorporated in402, and block406 simply replaces the original narrative with the stored narrative for that time card entry description.
• FIG.5 describes one embodiment of the operation of the machine learning model for a line of theinvoice501. First, the test of the narrative is parsed into a table of words502. This could use traditional parsing techniques of searching for delimiters such as spaces, periods, tabs, commas, and other punctuation types. The strings between the delimiters are stored in a table. In some descriptions, this is called tokenization
• The words in the table are next analyzed to convert the word into its stem by removing parts of speech that are attached to the word, such as plurality503. Some descriptions call finding this process lemmatization—figuring out the most basic form or lemma of each word in the sentence. For some models, the word is categorized as a noun, verb, adverb, adjective, etc. Steps502,503,504 are often referred to as natural language processing.
• Next, each word stem is looked up in the model's database of stems505, and the corresponding weight of the word in the database is copied into the narrative's table of words. In a simple model, the weights are averaged to create a confidence score506. More complicated models could use other statistical methods to determine the confidence score. For instance, verbs could be given more weight than adverbs, or a mean could be used rather than an average.
• Once the confidence score has been determined, the algorithm model backtracks to format the narrative text according to the impact each word has on theconfidence score507. This is done by looking at the table to see the relative weight of each word. The highest weighted words are identified, and the original text of the narrative is searched for each of these highest weighted words. When the word is found, it is converted to bold format, and the font is increased by a factor.
• The confidence score and the formatted narrative string are then returned508.
• Because of the complexities of machine learning algorithms, special purpose computing may be needed to build and execute the machine learning model described herein.FIG.6 shows one such embodiment. The user views the user interface described here on a personal computing device such as a personal computer, laptop, tablet, smartphone, monitor, orsimilar device601. Thepersonal computing device601 communicates through anetwork602 such as the Internet, a local area network, or perhaps through a direct interface to theserver603. Theserver603 is a high performance, multi- core computing device withsignificant storage facilities604 in order to store the training data for the model. Since this training data is continuously updated through the present inventions, this data must be kept online and accessible so that it can be updated. In addition, the real-time editing of the model as the user provides feedback to themachine learning model409 requires significant processing power to rebuild the model as feedback is received.
• Theserver603 is a high performance computing machine electrically connected to thenetwork602 and to thestorage facilities604.
• While the above described embodiment involves machine learning confidence scores, the algorithm could be used with confidence scores using other software techniques. For instance, it is envisioned that a heuristically based algorithm could also be used to analyze the narrative field, and that the word cloud could be formed by inserting formatting into the narrative field at each step in the heuristical algorithm.
• The foregoing devices and operations, including their implementation, will be familiar to, and understood by, those having ordinary skill in the art. This algorithm is necessarily rooted in computer technology in order to overcome the problem of displaying machine learning reasoning with a simple interface in order to receive user feedback in the machine's determination of a confidence score.
• The above description of the embodiments, alternative embodiments, and specific examples, are given by way of illustration and should not be viewed as limiting. Further, many changes and modifications within the scope of the present embodiments may be made without departing from the spirit thereof, and the present invention includes such changes and modifications.

Claims (20)

1. A method of comprising:

selecting an invoice for review by a computer, said computer electrically connected to a display screen; and

looping, by the computer, through each line of the invoice:

splitting the line into fields, said fields including a narrative field;

processing the narrative field through a machine learning model to derive a machine learning confidence score, wherein the processing by the machine learning model comprises:

converting the narrative field into word stems that are used by the machine learning model to calculate the machine learning confidence score,

searching the narrative field for the word stems that comprised a highest impact in the machine learning confidence score, and

adding formatting instructions to at least one word associated with the word stems that comprised the highest impact on the machine learning confidence score;

displaying on the display screen the fields, including the narrative field with the formatting instructions to indicate reasoning used by the machine learning model in a determination of the machine learning confidence score; and

displaying a variable icon on the display screen to graphically indicate the machine learning confidence score, wherein the variable icon displays a different number of items depending upon a magnitude of the machine learning confidence score.

2. The method ofclaim 1 wherein the invoice is a legal bill.

3. The method ofclaim 2 further comprising reviewing legal spending of a lawyer as described in the invoice.

4. The method ofclaim 3 wherein the reviewing includes observing the displayed narrative field and the variable icon.

5. The method ofclaim 3 wherein the review is conducted by a user.

6. The method ofclaim 1 wherein the invoice is a court reporter invoice.

7. The method ofclaim 1 wherein the invoice is an expert witness invoice.

8. The method ofclaim 1 wherein the narrative field with the formatting instructions is only displayed on the display screen when a mouse location is over an icon.

9. The method ofclaim 1 wherein a bold instruction is one of the formatting instructions.

10. The method ofclaim 1 wherein an instruction to increase font size is one of the formatting instructions.

11. The method ofclaim 1 wherein the variable icon displays three rays off of a lightbulb based on one range of the machine learning confidence score, and one ray off of the lightbulb based on another range of the machine learning confidence score.

12. A computer executing non-transitory machine readable media programmed to:

select an invoice for review by the computer, said computer electrically connected to a display screen; and

loop through each line of the invoice:

split the line into fields, said fields including a narrative field;

process the narrative field through a machine learning model to derive a machine learning confidence score, wherein the machine learning model converts the narrative field into word stems that are used by the machine learning model to calculate the machine learning confidence score, then searches the narrative field for the word stems that comprised a highest impact in the machine learning confidence score, and adds formatting instructions to at least one word associated with the word stems that comprised the highest impact on the machine learning confidence score;

display on the display screen the fields, including the narrative field with the formatting instructions to indicate reasoning used by the machine learning model in a determination of the machine learning confidence score; and

display a variable icon on the display screen to graphically indicate the machine learning confidence score, wherein the variable icon displays a different number of items depending upon a magnitude of the machine learning confidence score.

13. The computer ofclaim 12 wherein the invoice is a legal bill.

14. The computer ofclaim 13 further programmed to review legal spending of a lawyer as described in the invoice.

15. The computer ofclaim 12 wherein the invoice is a court reporter invoice.

16. The computer ofclaim 12 wherein the invoice is an expert witness invoice.

17. The computer ofclaim 12 wherein the narrative field with the formatting instructions is only displayed on the display screen when a mouse location is over an icon.

18. The computer ofclaim 12 wherein a bold instruction is one of the formatting instructions.

19. The computer ofclaim 12 wherein an instruction to increase font size is one of the formatting instructions.

20. The computer ofclaim 12 wherein the variable icon displays three rays off of a lightbulb based on one range of the machine learning confidence score, and one ray off of the lightbulb based on another range of the machine learning confidence score.

Images