Computer-implemented method and computer-readable medium for optimizing an order description for automatically assigning a service offer to the optimized order description

Abstract

A computer-implemented method for optimizing a job description (10) for automatically assigning a service offer (11) to the optimized job description (10), executed in a data processing system, the method comprising the steps of: receiving an initial job description (10) via a user interface of the data processing system, wherein a job description (10) comprises at least one specification of required work and / or requirements (V1); receiving multiple service offers (11) in the data processing system, wherein a service offer (11) comprises at least competencies and / or availabilities of a service provider (V2); determining comparison values ​​(1, 2, 3, ...), m) for the service offers (11) with the order description (10) by executing a first machine learning model (20), wherein the first machine learning model (20) receives the order description (10) and the service offers (11) as inputs, analyzes the context of the order description (10) and the respective service offer (11) for each of the service offers (11) and determines a respective comparison value (1, 2, 3, ..., m) based on a matching context between the order description (10) and the service offer (11) (V3) (...).

Description

[0001] The invention relates to a computer-implemented method and a computer-readable medium for optimizing an order description for automatically assigning a service offer to the optimized order description.

[0002] The following definitions, descriptions and statements retain their respective meaning for and apply to the entire disclosed subject matter of the invention.

[0003] Systems and processes for personnel recruitment, also known as recruitment, are known from the state of the art. Recruitment refers to the entire process of identifying, searching, selecting, pre-screening, and interviewing applicants for permanent or temporary positions within a company. Staffing also includes the process associated with selecting individuals for unpaid tasks. However, implementing recruitment requires a large amount of resources, such as managers, human resources professionals and recruitment specialists, public bodies, commercial recruitment agencies, or specialized personnel consulting firms. While internet-based technologies are known to improve aspects of recruitment, they still require a large amount of resources.

[0004] DE 10 2023 206 759 A1 discloses a computer-implemented method, a computer program, and a system for mediating between applicants and service providers within a company. The problem is that the number of hits determined there and the service providers' service descriptions are not used for the mediation to further optimize the order description and thus increase the probability of placement.

[0005] US 2022 / 0 129 856 A1 discloses a method and device for matching data using artificial intelligence. Based on a resume and a job profile, resume feature data and profile feature data are determined, and a knowledge graph path from the resume to the job profile is created and evaluated using a classification model. The path is then compared to the resume feature data of the first instance and the profile feature data of the second instance to determine a matching result between the first instance and the second instance.

[0006] Further state-of-the-art information is disclosed in KOINNOmagazin 02 / 23: Artificial Intelligence and Procurement Procedures. January 16, 2024. URL: https: / / www.koinno-bmwk.de / koinno / aktuelles / detail / kuenstliche-intelligenz-undvergabeverfahren / , LAWNICZAK, L.: AI in Procurement Law - Potentials and Risks in Comparison. May 22, 2024. URL: https: / / www.ibau.de / akademie / wissenswertes / ki-vergaberecht / , and SONG, K. [et al.]: MPNet: Masked and Permuted Pre-training for Language Understanding. November 2, 2020. URL: https: / / arxiv.org / pdf / 2004.09297.

[0007] The object of the invention was to find out how hit rates in order placement between client and service provider can be optimized for an automatic assignment of the service provider to an order description of the client.

[0008] The claimed subject matter each solves this problem. Advantageous embodiments of the invention emerge from the definitions, the dependent claims, the drawings, and the description of preferred embodiments.

[0009] According to one aspect, the invention provides a computer-implemented method for optimizing an order description for automatically assigning a service offer to the optimized order description. The method is executed in a data processing system.

[0010] The data processing system can be a company's IT platform, such as a cloud computing platform. This can be used to find internal experts and service providers. In this context, the data processing system is also called a matchmaking platform. This process allows a client, for example, to leverage the expertise of proven experts within their company. The client can, for example, enter their requirements as an order description in the data processing system and, upon completion of the process, receive offers from suitable service providers—these are the service offers. The commissioning or approval for a suitable service provider then takes place automatically within the data processing system.The process aims to optimize a contract description so that a provider's service offer can be automatically and efficiently assigned to that contract in the data processing system. For example, an algorithm or software analyzes or improves parameters or characteristics of a contract to enable efficient and targeted allocation of services that meet the contract requirements. The process can be used in various industries, such as logistics, where contracts are often awarded to external service providers, or in IT services, where the specific requirements of a project must be precisely defined in order to find the right offer. Thus, the process serves the purpose, among other things, of reducing the amount of resources used in contract placement within a company.Managers, human resources professionals and recruitment specialists, public bodies, commercial recruitment agencies or specialized personnel consulting companies can save money on contract placement.

[0011] In one step of the method, an initial job description is obtained via a user interface of the data processing system. The user interface can be an input device, for example, a terminal, a prompt input of application software, or a mobile device, for example, a notebook or a smartphone. A job description includes at least a specification of required work and / or requirements. For example, the client creates the job description themselves, and the job description thus created is received.

[0012] A job description refers to the detailed description or specification of a job or project. A job description typically includes information such as the scope of work required, the objectives of the job, specific requirements such as skills, human and / or machine resources, or timeframes, and often also criteria for measuring success. In the computer-implemented method disclosed here, the job description is digitized and / or structured so that it can be automatically analyzed and processed. The job description serves as the basis for matching with suitable service providers.

[0013] In a further step of the process, several service offers are received in the data processing system. A service offer includes at least the competencies and / or availability of a service provider. For example, service providers store descriptions of their services, expertise, and / or references in the data processing system. For example, service providers enter their service descriptions on the platform.

[0014] A service offering describes the portfolio or range of services that a provider or service provider can provide. The service offering can include information about the service provider's competencies, capacities, prices, and availability. The service offering is often presented in a format that allows it to be effectively matched with the requirements of the order description. In the computer-implemented method disclosed here, the service offerings are digitized and / or structured so that they can be automatically matched with relevant order descriptions.

[0015] In a further step of the process, comparison values ​​for the service offers with the order description are determined. The comparison values ​​are also called matching scores or hit rates. A higher hit rate indicates a higher match between the order description and the service offer. For this purpose, a first machine learning model is executed. The first machine learning model can be executed on a data processing unit of the data processing system, for example, on a hardware unit comprising CPU and / or GPU cores. The first machine learning model receives the order description and the service offers as input values. For each of the service offers, the first machine learning model analyzes the context of the order description and the respective service offer.This allows the first machine learning model to more deeply analyze and understand the context of order descriptions and service provider profiles, leading to more precise matches. Based on a matching context between the order description and the service offer, the first machine learning model determines a respective match score.

[0016] In a further step, the order description and those service offers whose respective comparison values ​​determined by the first machine learning model exceed a predefined threshold are entered into a generative second machine learning model.

[0017] A generative machine learning model is trained to generate new data that reflects certain properties or patterns of the training data. Generative machine learning models learn the underlying probability distribution of the data and can generate new instances from it, for example, image generation based on a text description, as is known from the prior art, or question generation based on a task description, as in the method disclosed here.

[0018] The second machine learning model can be controlled via the user interface with prompts. Prompts are data inputs. The data inputs describe the task that the second machine learning model should perform. The data inputs can be text inputs, for example. Using prompts to control the second machine learning model corresponds to intent-based outcome specification, in which the user tells the second machine learning model what they want, not how the second machine learning model should do it.

[0019] As can be seen from the further method steps, the invention utilizes the combination of the first machine learning model, which determines the matching values, and the generative second machine learning model, which performs semantic clarifications based on the matching values ​​and thus optimizes the order description. Generative models alone do not provide precise hit rates. The first machine learning model is not capable of performing semantic clarifications. The symbiosis of the first and second machine learning models disclosed here, which is a true combination, enables iterative improvement of the order content by generating clarifying queries in real time. This leads to a significant improvement in the precision of matching between clients and service providers, thus resulting in an increased hit rate, an increased matching probability, and resource savings.

[0020] The control of the second machine learning model and the combination of the first machine learning model with the second machine learning model falls within the sphere of the technical expert. The associated interaction of two machine learning models enables an iterative improvement of the order description by generating clarifying queries in real time, as disclosed below. This leads to a significant improvement in the precision of matching between clients and service providers and thus results in increased matching values. The combination of the first machine learning model with the second machine learning model is a technical implementation and, as such, contributes to the technical nature of the method. The combination of the first machine learning model with the second machine learning model is a specific technical implementation of a matching process, which results in a technical improvement in terms of saving resources.

[0021] In a further step, prompt engineering instructs the second machine learning model to generate questions based on the answers to the questions, increasing the comparison scores of the service offers entered into the second machine learning model with the order description. The prompts are pre-formulated by system administrators and contain variable components. For example, a kind of basic framework for the prompt can be specified: "Analyze the following job description and determine which field of expertise it fits. Then act in the role of a recruiter with in-depth understanding of this field. I will provide my input in the following JavaScript Object Notation (JSON) structure: { "Job Description" : " <auftragsbeschreibung>", "Service Offer 1": " <leistungsbeschreibung>", "Service Offer2": " <leistungsbeschreibung>", "Service Offer3": " <leistungsbeschreibung>", "Number of queries": " <anzahlrueckfragen>",}

[0022] Your task now, in your role as a recruiter with a deep understanding of the area of ​​expertise you've identified, is to generate the desired number of clarifying questions from the job description and the service offerings. Please ask the questions in simple language and include potential examples in parentheses after the question.

[0023] The questions generated by the second machine learning model are issued to a user. As can be seen from the prompt mentioned above, prompt engineering refers to the process, strategy, and tactics used to design and construct this prompt. The focus is on shaping the prompt to most effectively guide the behavior or response of the second machine learning model.

[0024] These prompts must be carefully crafted by the system administrator to generate specific questions to ensure that the second machine learning model delivers accurate and relevant results. This requires expertise and professional skill on the part of the system administrator.

[0025] According to another aspect, specific prompting is used to formulate questions in a domain-specific manner. For example, in the area of ​​research and development, the questions are technically detailed, while in other areas, they are more general. For example, this can be used to assign the second machine learning model a role it should assume, such as mediation expert for research and development projects.

[0026] In a further step, the user's questions and corresponding answers are entered into the second machine learning model, which corresponds to an interaction of the second machine learning model, for example, through user feedback. The second machine learning model generates the optimized order description based on the answers. Based on the user's interactions with the queries generated by the second machine learning model and the corresponding adjustments (i.e., optimizations) to the order description, the matching values ​​(i.e., the hit rates) are continuously updated and optimized to maximize the probability of successful matchmaking between orders and services and further save resources.

[0027] In a further step, the previous steps are iterated to increase the adjustment values ​​determined by the first machine learning model, whereby the optimized job description is obtained as the initial job description.

[0028] According to a further aspect, the relevance of technical terms in the order description and the respective service offer is determined to determine matching scores. This relevance is combined with the first machine learning model's analysis of the context of the order description and the respective service offer. This combination is also called an overall score. The respective matching score is determined from this combination. This further improves the accuracy of the matching.

[0029] According to the invention, the relevance of technical terms in the order description and the respective service offer is determined based on the term frequency-inverse document frequency (TF-IDF) measure to determine matching values. In the method disclosed here, this measure is used to weight the relevance or importance of specific terms, such as key terms, in the order descriptions and service offers. This further improves the accuracy of the matching.

[0030] The term frequency-inverse document frequency TF-IDF measure is widely used in information retrieval and characterizes the importance of a word for a document in a collection or corpus, adjusted for the fact that some words generally appear more frequently.

[0031] The occurrence frequency indicates how often the term occurs in m documents. For example, if the document is the sentence "The red car stops at the red light.", then To avoid distorting the results in long documents, it is possible to normalize the absolute occurrence frequency. To do this, the number of occurrences of a term in the document is divided by the maximum frequency of a term in and one obtains the relative occurrence frequency The inverse document frequency of a term does not depend on the individual document, but on the document corpus, i.e., the total set of all documents in the retrieval scenario: Here, is the number of documents in the corpus and is the number of documents that contain the term.

[0032] With the weighting of a word calculated in this way with respect to the document in which it is contained, documents can be better arranged in the result list as search hits of a word-based search than would be possible, for example, using term frequency alone, see the corresponding Wikipedia article, CC-BY-SA licensed, accessed on May 15, 2024 .

[0033] According to the invention, the first machine learning model is pre-trained on a dataset comprising text according to the masked and permuted MPNet method. During the implementation of the method disclosed here, the first machine learning model is retrained on received order descriptions and service offers.

[0034] The MPNet method is disclosed, for example, in "MPNet: Masked and Permuted Pre-training for Language Understanding", Kaitao Song et al., arXiv:2004.09297v2 [cs.CL] 2 Nov 2020, section 2. Using the MPNet method, the first machine learning model is particularly effective in understanding the context within texts because it uses both masking and permutation during training. The dataset on which the first machine learning model is or is pre-trained, also called pre-training, can include, for example, Wikipedia, BooksCorpus https: / / hug-gingface.co / datasets / bookcorpus, OpenWebText https: / / cdn.openai.com / better-language-models / language_models_are_unsupervised_multitask_learners.pdf, CommonCrawl-News dataset arXiv:1907.11692v1 or Stories arXiv:1806.02847 or any combination of the aforementioned datasets.Through retraining, also called fine-tuning, on process-specific data, such as order descriptions, service offers, or historical data related to the process's execution, the first machine learning model becomes specialized. This specialization enables the first machine learning model to recognize finer nuances in the requirements and offers.

[0035] Implementing or applying the MPNet method and retraining the first machine learning model requires a specialist.

[0036] According to one aspect, the aforementioned overall score, i.e., the combination of the relevance of technical terms in the order description and the respective service offer with the first machine learning model's analysis of the context of the order description and the respective service offer, is calculated as follows: Let "mpnet score" be the matching value of the first machine learning model obtained from the pre-trained and fine-tuned MPNet procedure. Let "tf-idf score" be the relevance of technical terms in the order description and the respective service offer. The overall score is then obtained as follows: overall−score=(mpnet−score*K+tf−idf−score*(1−K)), where K is an adjustable adjustment factor that can be used to weight the importance of specific terms in the descriptions, which further improves the accuracy of the matching.

[0037] A combination of the TF-IDF measure with the MPNet method, as claimed by the references in the patent claims, is particularly advantageous for determining the overall score.

[0038] According to a further aspect, user feedback is received via the user interface for evaluating the matching values ​​determined by the first machine learning model. The first machine learning model is retrained using the feedback. Using user feedback, the first machine learning model is optimized. This allows the relevance and accuracy of the matching values ​​to be refined. User feedback can, for example, relate to whether the received match is partially a match, not a match at all, a perfect match, a match largely a match, or a reasonable match. This corresponds to training with human feedback.

[0039] According to another aspect, the second machine learning model is a large-language model (LLM). Large-language models can be controlled via prompt engineering. A large-language model can generate questions to increase the correlation values ​​of the service offers entered into the second machine learning model with the order description based on the answers to the questions. The second machine learning model can, for example, be a generative pre-trained transformer, such as gpt-3.5, gpt-4, gpt-4o, or an LLaMA language model, such as LLaMA 3.

[0040] The specific combination and interaction of the first machine learning model, which is pre-trained according to the MPNet method and retrained on received order descriptions and service offers during the implementation of the method disclosed here, with the second machine learning model in the form of a large-language model is a specific technical implementation of a matching process that brings about a technical improvement in terms of saving resources.

[0041] According to a further aspect, the service offers with the respective comparison values ​​are displayed via the user interface for the order description. This corresponds to a summary analysis and evaluation of the offers for the client, thus supporting informed decision-making.

[0042] According to a further aspect, a service offer is optimized for the automatic assignment of a job description to the optimized service offer. This corresponds to a kind of role reversal between the job description and the service offer with regard to the method disclosed here. The method disclosed here can therefore be used for both job descriptions and service offers. The service provider, who is an expert in their field and would like to offer their expertise as a service, can thus find and receive job descriptions that match their service offer in a resource-efficient manner.

[0043] According to a further aspect, after the step of determining the respective comparison value, a user query is conducted as to whether questions to refine the order description should be answered. If no questions are to be answered, the order description is published in the data processing system.

[0044] According to a further aspect, user feedback regarding the first threshold is actively incorporated into the training of the first machine learning model, which enables continuous improvement of the first machine learning model. User feedback helps refine the relevance and accuracy of the comparison values.

[0045] In another aspect, the invention provides a computer-readable medium comprising instructions that, when executed by a computer, cause the computer to perform the steps of the method claims disclosed herein. The computer-readable medium may be a computer-readable storage medium that can control a general-purpose computer, a special-purpose computer, or other programmable data processing device to perform the method disclosed herein. The instructions are, for example, program instructions.

[0046] The computer-readable medium may be a tangible device capable of storing instructions for use by an instruction execution device. The computer-readable storage medium may be, for example, an electronic storage device, a magnetic storage device, an optical storage device, an electromagnetic storage device, a semiconductor storage device, or any suitable combination thereof.A non-exhaustive list of more specific examples of the computer-readable storage medium includes the following: a portable computer diskette, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), static random access memory (SRAM), portable compact disc read-only memory (CD-ROM), a digital versatile disk (DVD), a memory stick, a floppy disk, a mechanically encoded device such as punched cards or raised structures in a groove with instructions recorded thereon, and any suitable combination of the foregoing. A computer-readable storage medium, as used herein, may, but need not, be construed as transient signals in and of themselves, such as radio waves or other freely propagating electromagnetic waves, electromagnetic waves propagating through a waveguide or other transmission media (e.g., a radio frequency (RF) signal).B. light pulses passing through a fiber optic cable) or electrical signals transmitted through a wire.

[0047] According to a further aspect, the first and / or second machine learning model issues recommendations for competency enhancements to service providers, for example based on trend analyses and demand patterns within the data processing system, in order to increase market opportunities.

[0048] The invention is illustrated in the following embodiment of Fig. 1 illustrates this.

[0049] Fig. 1 shows a computer-implemented method for optimizing an initial order description 10 for automatically assigning a service offer 11 to the optimized order description 13. The method is carried out in a data processing system, for example in a data processing system of a company, for example in a cloud-based data processing system.

[0050] In a step V1, an initial order description 10 is received via a user interface of the data processing system. The user interface can be, for example, a mobile device, such as a notebook.

[0051] In a step V2, several service offers 11 are received in the data processing system, for example via the user interface, via which service providers, for example internal company service providers, are also connected.

[0052] In a step V3, comparison values ​​1, 2, 3, ..., m for the service offers 11 with the order description 10 are determined by executing a first machine learning model 20. To determine the comparison values ​​1, 2, 3, ..., m, the first machine learning model 20 can, for example, determine a relevance of technical terms in the order description 10 and the respective service offer 11 based on the term frequency-inverse document frequency TF-IDF measure. In the embodiment of the Fig. 1, the first machine learning model 20 is also pre-trained according to the MPNet method and is retrained on data related to order descriptions and service offers. The first machine learning model 20 receives the order description 10 and the service offers 11 as inputs. In the embodiment of the Fig. 1, the TF-IDF measure is combined with the MPNet method, resulting in the matching values ​​1, 2, 3, ..., m: The first machine learning model 20 determines the relevance of the technical terms in the order description 10 and the respective service offer 11 based on the TD-IDF measure. For each of the service offers 11, the first machine learning model 20 further analyzes the context of the order description 10 and the respective service offer 11. Pre-training using masking and permutation makes the first machine learning model 20 particularly suitable for analyzing contexts. Based on the relevance of the technical terms in the order description 10 and the respective service offer 11, and based on a matching context between the order description 10 and the service offer 11, a respective matching value 1, 2, 3, ... is then determined., m is determined, i.e. matching score 1 for matching between order description and a first service offer 11, matching score 2 for matching between order description and a second service offer 11, matching score 3 for matching between order description and a third service offer 11, etc.

[0053] In a step V4, a user query is conducted, for example via the user interface, as to whether questions should be answered to refine the order description 10. If not, in a step V5, the order description 10 is published in the data processing system, for example, displayed on a platform visible to other users.

[0054] Otherwise, in a step V6, the order description 10 and those service offers 11 whose respective comparison values ​​1, 2, 3, ..., m determined by the first machine learning model 20 exceed a predeterminable threshold value 12 are entered into a generative second machine learning model 30. The second machine learning model 30 can be controlled with prompts.

[0055] In a step V7, prompt engineering 40 instructs the second machine learning model 30 to generate questions 50 in order to increase the comparison values ​​1, 2, 3, ..., m of the service offers 11 entered into the second machine learning model 30 with the order description 10 based on answers 51 to the questions 50. The questions 50 are output to the user, for example, via the user interface.

[0056] In a step V8, the user's questions 50 and the corresponding answers 51 are entered into the second machine learning model 30. In a step V9, the second machine learning model 30 generates the optimized order description 13 based on the answers 51.

[0057] In a step V10, the preceding steps are iterated to increase the adjustment values ​​1, 2, 3, ..., m determined by the first machine learning model 20, whereby the optimized job description 13 is obtained as the initial job description 10. Reference symbol 10 initial order description 11 Range of services 13 optimized order description 1, 2, 3, ..., m adjustment value 12 Threshold 20 first machine learning model 30 second machine learning model 40 Prompt Engineering 50 questions 51 answers< / anzahlrueckfragen> < / leistungsbeschreibung> < / leistungsbeschreibung> < / leistungsbeschreibung> < / auftragsbeschreibung>

Claims

[1] Computer-implemented method for optimising an initial order description (10) for automatically assigning a service offer (11) to the optimised order description (13), carried out in a data processing system, the method comprising the steps: • receiving the initial order description (10) via a user interface of the data processing system, wherein an order description (10, 13) comprises at least one specification of required work and / or requirements (V1); • Receiving a plurality of service offers (11) in the data processing system, wherein a service offer (11) comprises at least competencies and / or availabilities of a service provider (V2); • Determining comparison values ​​(1, 2, 3, ..., m) for the service offers (11) with the order description (10) by executing a first machine learning model (20), wherein the first machine learning model (20) receives the order description (10) and the service offers (11) as inputs, analyzes the context of the order description (10) and the respective service offer (11) for each of the service offers (11) and determines a respective comparison value (1, 2, 3, ..., m) based on a matching context between the order description (10) and the service offer (11) (V3), wherein to determine the comparison values ​​(1, 2, 3, ..., m) a relevance of technical terms in the order description (10) and the respective service offer (11) is determined based on the term frequency-inverse document frequency TF-IDF measure; • entering the order description (10) and those service offers (11) whose respective comparison values ​​(1, 2, 3, ..., m) determined by the first machine learning model (20) exceed a predeterminable threshold value (12) into a generative second machine learning model (30), wherein the second machine learning model (30) is controllable with prompts (V6); • by prompt engineering (40) instructing the second machine learning model (30) to generate questions (50) (V7) in order to increase the comparison values ​​(1, 2, 3, ..., m) of the service offers (11) entered into the second machine learning model (30) with the order description (10) based on answers (51) to the questions (50), and outputting the questions (50) to a user; • Entering the questions (50) and the associated answers (51) of the user into the second machine learning model (30) (V8), wherein the second machine learning model (30) generates the optimized order description (13) based on the answers (51) (V9); • wherein the first machine learning model (20) is pre-trained on a data set comprising texts according to the masked and permuted MPNet method and is retrained on received order descriptions (10, 13) and service offers (11) during the execution of the method; • Iteration of the preceding steps to increase the adjustment values ​​(1, 2, 3, ..., m) determined by the first machine learning model (20), wherein the optimized job description (13) is obtained as the initial job description (10) (V10). [2] Method according to claim 1, wherein, in order to determine comparison values ​​(1, 2, 3, ..., m), a relevance of technical terms in the order description (10) and the respective service offer (11) is determined and the relevance is combined with the analysis of the context of the order description (10, 13) and the respective service offer (11) and the respective comparison value (1, 2, 3, ..., m) is determined from this combination. [3] Method according to one of the preceding claims, wherein feedback from the user is received via the user interface for an evaluation of the adjustment values ​​(1, 2, 3, ..., m) determined by the first machine learning model (20) and the first machine learning model (20) is retrained with the feedback. [4] Method according to one of the preceding claims, wherein the second machine learning model (30) is a large-language-model LLM. [5] Method according to one of the preceding claims, wherein for the optimized order description (13) the service offers (11) with the respectively determined comparison values ​​(1, 2, 3, ..., m) are output via the user interface. [6] Method according to one of the preceding claims, wherein a service offer (11) is optimized for an automatic assignment of an order description (10, 13) to the optimized service offer (11). [7] Method according to one of the preceding claims, wherein after the step (V3) of determining the respective comparison value, a user query is carried out as to whether questions for refining the order description (10) are to be answered (V4) and, in the event that no questions are to be answered, the order description (10) is published in the data processing system (V5). [8] A computer-readable medium comprising instructions which, when executed by a computer, cause the computer to carry out the steps of the method according to any one of the preceding claims.

Images