A Case for Business Process-Speciﬁc Foundation Models Yara Rizk

2025-04-30 0 0 295.39KB 9 页 10玖币

侵权投诉

A Case for Business Process-Speciﬁc Foundation

Models

Yara Rizk

IBM Research

yara.rizk@ibm.com

Praveen Venkateswaran

IBM Research

praveen.venkateswaran@ibm.com

Vatche Isahagian

IBM Research

vatchei@ibm.com

Vinod Muthusamy

IBM Research

vmuthus@us.ibm.com

Abstract

The inception of large language models has helped advance state-of-the-art per-

formance on numerous natural language tasks. This has also opened the door

for the development of foundation models for other domains and data modalities

such as images, code, and music. In this paper, we argue that business process

data representations have unique characteristics that warrant the development of a

new class of foundation models to handle tasks like process mining, optimization,

and decision making. These models should also tackle the unique challenges

of applying AI to business processes which include data scarcity, multi-modal

representations, domain speciﬁc terminology, and privacy concerns.

1 Introduction

Artiﬁcial intelligence, especially since the emergence of deep learning, has disrupted many areas

of our lives from personal assistants like Alexa [Schneider, 2020] to autonomous driving [Bernhart

and Winterhoff, 2016]. It has also been a disruptive force for businesses

; deep learning is estimated

to provide between $3.5 trillion and $5.8 trillion of annual value Chui et al. [2018] and can be the

difference between companies’ rise or demise.

In enterprise settings, business processes provide a structured framework for work. They deﬁne

tasks, and identify their executors while capturing dependencies and providing logging and tracking

capabilities. They also capture company policies and compliance with regulations. With many

enterprises relying on the business process management paradigm to standardize their work, process

management tools grew to a $11.84 billion industry and is projected to grow to $26 billion in 2028

However, the existing landscape of work has been rapidly changing, requiring companies to move

from their static business process practices to more agile and automated methods due to increased

supply chain disruptions and skill shortages from the recent pandemic. Thus, companies are making

signiﬁcant investments to adopt AI-driven tools for tasks like process prediction, visualization,

translation, etc. [McKendrick, 2021], evidenced by the $1+ million investments made by companies

and the projected $3.2+ trillion business value produced by AI tech gar [2018]. Foundation models’

recent success presents an opportunity to improve business process automation and management.

1https://www.gartner.com/smarterwithgartner/the-disruptive-power-of-artiﬁcial-intelligence

https://www.marketwatch.com/press-release/business-process-management-market-size-growth-with-top-

leading-players-growth-key-factors-global-trends-industry-share-and-forecast-2022-2031-2022-08-18

https://www.gartner.com/en/newsroom/press-releases/2021-09-29-gartner-ﬁnds-33-percent-of-

technology-providers-plan-to-invest-1-million-or-more-in-ai-within-two-years

Preprint. Under review..

arXiv:2210.14739v2 [cs.AI] 30 Nov 2022

Record loan

application

Verify

employment

Review credit

report

Request credit

report

Review title

report

Perform title

Review loan

application

Send rejection

Send approval

Figure 1: Example of a mortgage loan application process (Source: Chakraborti et al. [2020b])

Similar to natural language, images, or code snippets, business processes are yet another information

representation paradigm. However, the unique and particular nature of process features and modalities

can render existing foundation models inadequate to accurately understand and reason over them.

Hence, developing successful foundation models for business process decision making requires

research efforts to treat process data in a holistic manner instead of separate, independent modalities.

In this paper, we propose an approach to creating foundation models that factor in the complexity

of process data. We also discuss some of the challenges of creating foundation models for business

processes and the risks and opportunities of foundation models’ emergent behavior. First, however,

we provide an overview of business processes, their unique properties and the tasks

that may be best

suited for foundation models.

2 Background

2.1 Business Process Management

Abusiness process is a collection of ordered tasks, followed by a business to produce a product

or a service Weske [2012]. Figure 1 shows the example of a mortgage application process where

every application must go through the same steps before a decision is made. This allows mortgage

lenders to structure their process, improve consistency across loan ofﬁcers and track the execution of

the process for accountability, auditing and improving the provided service. A graphical notation,

known as business process model and notation (BPMN) Grosskopf et al. [2009], is generally used to

represent such processes, capturing the relationship between tasks (rectangular boxes with rounded

edges) that must be completed by employee roles within an organization, events (circles) that can

trigger processes or speciﬁc tasks within them, and decision points (diamonds) that allow paths within

the process ﬂow to merge or diverge. Swim lanes are usually deﬁned to place speciﬁc tasks within

the scope of an employee role or department. A trace is an execution of a process; each process can

produce many distinct traces when executed depending on input events and other factors.

Business process management consists of many problems related to the modeling or design, execution

and governance of processes. Process mining or discovery analyzes event data to identify and derive

processes from raw, unstructured data Van Der Aalst [2012]. Ideally, process mining should produce

a BPMN or similar representation for the discovered process. Process optimization or re-engineering

looks to improve existing processes Arlbjørn and Haug [2010]. This requires making changes

to the process representation while maintaining the properties that characterize a valid process.

Conformance checking veriﬁes that the “as-is” process (i.e., how the process is being executed in

reality) does not deviate from the “to-be” process (i.e., how the process was theoretically designed

to be executed) Dunzer et al. [2019]. Task automation through robotic process automation looks to

create automation scripts that can programmatically execute tasks instead of humans Van der Aalst

et al. [2018], whereas automation mining programmatically identiﬁes the best tasks to automate

Geyer-Klingeberg et al. [2018].

2.2 Foundation Models

Foundation models, coined in Bommasani et al. [2021], refer to deep neural network models trained

on massive data and can be reused (with minimal modiﬁcations) for multiple downstream tasks.

Since processes also call individual nodes within a process a task, we will use “downstream tasks” to refer

to foundation model speciﬁc prediction tasks and “process tasks” to refer to tasks within a process.

文档加载中……请稍候！
如果长时间未打开，您也可以点击刷新试试。

下载文档到电脑，查找使用更方便

10 玖币 0人已下载

立即下载

摘要：

ACaseforBusinessProcess-SpecicFoundationModelsYaraRizkIBMResearchyara.rizk@ibm.comPraveenVenkateswaranIBMResearchpraveen.venkateswaran@ibm.comVatcheIsahagianIBMResearchvatchei@ibm.comVinodMuthusamyIBMResearchvmuthus@us.ibm.comAbstractTheinceptionoflargelanguagemodelshashelpedadvancestate-of-the-art...

展开>> 收起<<

A Case for Business Process-Speciﬁc Foundation Models Yara Rizk.pdf

共9页,预览2页

还剩页未读，继续阅读

声明：本站为文档C2C交易模式，即用户上传的文档直接被用户下载，本站只是中间服务平台，本站所有文档下载所得的收益归上传人(含作者)所有。玖贝云文库仅提供信息存储空间，仅对用户上传内容的表现方式做保护处理，对上载内容本身不做任何修改或编辑。若文档所含内容侵犯了您的版权或隐私，请立即通知玖贝云文库，我们立即给予删除！

A Case for Business Process-Speciﬁc Foundation Models Yara Rizk

相关推荐

开通VIP享超值会员特权

作者详情

相关内容

热门标签

举报选择: