[Paper Review] Online Similarity Learning with Feedback for Invoice Line Item Matching

Chandresh Kumar Maurya, Neelamadhav Gantayat+2 more|arXiv (Cornell University)|Jan 2, 2020

Data Stream Mining Techniques16 references4 citations

TL;DR

This paper proposes online similarity learning methods for invoice line item matching in procure-to-pay processes, using agent feedback to train ranking or classification models. By integrating product taxonomy and catalogs, the approach improves matching accuracy and efficiency over baselines, reducing manual effort in enterprise invoice processing.

ABSTRACT

The procure to pay process (P2P) in large enterprises is a back-end business process which deals with the procurement of products and services for enterprise operations. Procurement is done by issuing purchase orders to impaneled vendors and invoices submitted by vendors are paid after they go through a rigorous validation process. Agents orchestrating P2P process often encounter the problem of matching a product or service descriptions in the invoice to those in purchase order and verify if the ordered items are what have been supplied or serviced. For example, the description in the invoice and purchase order could be TRES 739mL CD KER Smooth and TRES 0.739L CD KER Smth which look different at word level but refer to the same item. In a typical P2P process, agents are asked to manually select the products which are similar before invoices are posted for payment. This step in the business process is manual, repetitive, cumbersome, and costly. Since descriptions are not well-formed sentences, we cannot apply existing semantic and syntactic text similarity approaches directly. In this paper, we present two approaches to solve the above problem using various types of available agent's recorded feedback data. If the agent's feedback is in the form of a relative ranking between descriptions, we use similarity ranking algorithm. If the agent's feedback is absolute such as match or no-match, we use classification similarity algorithm. We also present the threats to the validity of our approach and present a possible remedy making use of product taxonomy and catalog. We showcase the comparative effectiveness and efficiency of the proposed approaches over many benchmarks and real-world data sets.

Motivation & Objective

• To address the high cost and manual effort of invoice line item matching in enterprise procure-to-pay (P2P) processes.
• To develop a similarity learning framework that leverages agent feedback—either relative rankings or binary match/no-match labels—for improved matching accuracy.
• To incorporate product taxonomy and catalogs to resolve hierarchical and semantic mismatches in line item descriptions.
• To reduce reliance on rule-based or keyword-matching approaches that fail with OOV words, OCR errors, and syntactic variations.
• To demonstrate the effectiveness and efficiency of the proposed method on real-world and benchmark datasets.

Proposed method

• The method uses agent feedback in two forms: relative rankings (for similarity ranking) or binary match/no-match labels (for classification).
• A similarity ranking algorithm is trained using feedback to prioritize correct matches over incorrect ones in a latent embedding space.
• A classification similarity algorithm is applied when feedback is binary, learning to predict match/no-match outcomes.
• The approach integrates product taxonomy to detect hierarchical relationships (e.g., hyponymy), such as 'edible oil' being a category of 'coconut oil'.
• Product catalogs are used to extract and match product attributes and entities, improving semantic alignment between invoice and PO descriptions.
• A hybrid similarity measure based on Jaccard similarity is computed over merged, matched, and remaining tokens to produce a final similarity score.

Experimental results

Research questions

• RQ1 How can online similarity learning with agent feedback improve the accuracy of invoice line item matching in P2P processes?
• RQ2 What is the impact of incorporating product taxonomy and catalogs on resolving hierarchical and semantic mismatches in invoice descriptions?
• RQ3 How does the proposed method outperform rule-based or keyword-matching approaches in handling OOV words, OCR errors, and syntactic variations?
• RQ4 Can feedback in the form of relative rankings or binary labels be effectively leveraged to train robust similarity models in real-time?
• RQ5 What role does product taxonomy play in improving the matching of general or broad product categories across invoices and purchase orders?

Key findings

• The proposed method significantly reduces manual effort in invoice processing by automating line item matching using agent feedback and domain knowledge.
• Incorporating product taxonomy improves matching accuracy by resolving hierarchical relationships, such as matching 'edible oil' to 'coconut oil'.
• Using product catalogs enhances matching by aligning attributes and entities, even when descriptions vary in phrasing or include additional details like quantity.
• The similarity ranking and classification models outperform baseline approaches on real-world and benchmark datasets, demonstrating improved effectiveness and efficiency.
• The hybrid approach using Jaccard similarity over matched, common, and remaining tokens yields a robust and interpretable similarity score.
• The method effectively handles challenges like OCR errors, OOV words, and syntactic variations (e.g., '10 boxes of pencils' vs '10 pencil boxes').