To assist knowledge work in accounting use cases such as bookkeeping, this paper presents a pipeline for constructing and enriching an accounting knowledge graph from heterogeneous accounting resources. To show the feasibility of the approach, we applied the pipeline in a multi-group energy provider by employing real company data. A set of prototypical knowledge services was realized with the accounting knowledge graph as the basis, for example, the suggestion of similar accounting cases to the accountant. For training decision trees to predict accounts, our results suggest that using semantically enriched data from the knowledge graph leads to better results compared to not using semantically enriched data.
The work of accountants is characterized by searching relevant information which is distributed in diferent company sources. One typical task is bookkeeping where besides verifying an invoice, decisions are made about the set of accounts and VAT rates that need to be selected for a particular accounting case. For solving such tasks, accountants typically tap into several data sources, for example, chart of accounts, accounting manuals and handbooks, or accounting policies. In multi-group companies, accounting departments are usually responsible for all subsidiary groups which means that they have to consider such documents for each group separately. Further common sources are personal notes, referenced documents on an invoice, historical accounting cases as references, additional attachments, or the knowledge of a colleague.
Tapping into this many data sources can lead to cases where the time needed to find the
relevant information is high, especially for non-standard cases. Jain and Woodcock [
Use cases around the Financial Industry Business Ontology (FIBO) [
Because the envisioned knowledge services require upfront queries over information that resides in the diferent introduced data sources, which have diferent formats as well such as PDF or CSV, and because the work context of an accountant is important but nowhere modeled and displayed in the task process explicitly, we decided to first construct an accounting knowledge graph and then to implement the knowledge services on top of it.
Figure 1 depicts the pipeline of constructing and enriching an accounting knowledge graph. The data basis was a CSV file containing 65k processes from the year 2019. It contained information such as the invoice issuer, invoice issuer number, booking area, the service description, or the amount of money. According to requirement a), we extracted at first all processes that deal with food and beverage scenarios. For this, two approaches were followed: On the one hand, with a keyword list, we mimicked the approach an accountant would use by searching the service descriptions of invoices for keywords such as "secretariat service" or "restaurant". On the other hand, we filtered processes which were booked on standard food and beverage accounts where food and beverage cases are booked exclusively. By merging the results, 1267 food and beverage processes have been detected which represent the first input for the knowledge graph.
To construct the first version of the knowledge graph (KG v1 in Figure 1), another CSV file
containing a set of booking statements for each process has been considered. Such booking
statements contain, among others, an account, a VAT rate identifier and the accordingly split
amount of money. Technology-wise, we used RML [
For the second version of the knowledge graph (KG v2 in Figure 1), the type of invoice issuer was refined as well as the service description. By means of keyword matching, triples to the invoice issuer have been automatically added which state whether the issuer is, for example, a restaurant, a hotel, or a caterer. In the same way, triples have been added to the service description to specify whether the invoice is, for example, about a breakfast, a business lunch, or a meeting service. For a full list of enrichments, we would like to refer to the ontology page. To reach KG v3 in Figure 1, we incorporated information about accounts which resides in multiple PDF handbooks for each company group separately. Accordingly, per account, information about the account ID, short- and long descriptions as well as the provenance in form of the accounting handbook were incorporated into the knowledge graph. Because the accounting handbooks were all structured identically starting with the account identifier per account, and by exploiting this structure with regular expressions, the handbook text could be processed automatically. To enable the suggestions of colleagues (Figure 1, KG v4), we firstly included all 1https://github.com/RMLio/rmlmapper-java 2By searching in ontology repositories such as https://lov.linkeddata.es/dataset/lov/ accountants as person resources with their contact data. For linking accountants to historical processes, historic process protocols have been analyzed to find out which accountants were involved in a process (also by employing regular expressions).
Because the goal was to embed knowledge services directly into the workspace of an accountant,
information is shown in form of a sidebar on the desktop. The screenshot in Figure 2 shows an
excerpt of the suggestions of similar accounts. In this way, an accountant can look into such
processes to get clues for processing the current accounting case. For calculating the similarity
between processes, we leveraged the semantically enriched information about types of invoice
issuer and services. An insight from interviews was that the similarity indicators needed to be
weighted diferently so that the booking area is more important than the invoice issuer and the
invoice issuer is more important than the type of service. Therefore, initially, a ratio of 4:2:1:1
was adopted (where the additional "1" represents the type of invoice issuer). Because during
enrichment accountants have been linked to historic processes, the similarity matrix could also
be used for suggesting colleagues. For account and VAT rate prediction, two approaches have
been followed: First, based on the similarity matrix, we suggest the accounts and VAT rates
of similar processes. Second, Table 1 summaries results for learning decision trees by using
diferent degrees of semantically enriched data: using non-enriched data and using semantically
enriched data with (I) the type of the invoice issuer, (II ) the type of the service, and (III) with the
combination of both. WEKA [
In this paper, we presented a pipeline for constructing and enriching an accounting knowledge graph for bookkeeping use cases. To enable reuse, we further proposed required bookkeeping vocabulary. Besides showing the feasibility of realizing knowledge services on top of such an accounting knowledge graph, first results indicate that using semantically enriched data leads to better results when learning decision trees to predict accounts. One lane for future work is the ofering of explanations in form of derived rules.
Acknowledgements: This work was funded by BMBF project SensAI (grantno. 01IW20007).