AI for retail

Practical tips while applying AI and ML to generate eCommerce product content

Commerce has evolved to eCommerce, and the ‘e’ stands for more than ‘electronic’. It’s more about ‘easy’ now – easier to find products, compare across brands, purchase them and get them delivered on time. An important factor that helps make informed purchase decisions with ease, is to have complete product information on the product page. But, ever wondered what happens behind the scenes to create this seamless and ‘easy’ experience?

Imagine if a retailer needs to sell a simple, black office chair. How would the retailer describe that black office chair? The description needs to be factually accurate, unique, have all features listed, use the right keywords, adhere to the company style, language, and brand guidelines… the list is exhaustive. Getting in all these details and more, manually, for one office chair, can be daunting. Imagine having to do this for millions of SKUs – every day!

Consequently, retailers are increasingly turning to technology-enabled artificial intelligence (AI) and machine learning (ML) to automate the process and generate product content at scale. In our experience, leading Fortune 500 businesses are witnessing an approx. 10% improvement in conversion and 12% lift in search traffic by generating automated product content. Here are some practical tips while applying AI / ML to generate eCommerce product content:

Stakeholder alignment and change management

Leveraging AI / ML to generate product content has several benefits. But its smooth implementation is subject to stakeholder buy-in. Therefore, align with key stakeholders across teams early on, on the goal and ROI from using AI-enabled processes. Set clear cadence and expectations on what can and cannot be delivered using AI, and ensure they align with overall business goals. This will also help in managing change effectively.

Relevant application –where it matters

Artificial intelligence and machine learning are popular buzzwords, but the key is to implement it where it can solve the business problem. Stay laser focused on the business need – for instance, if the need is to write creative product descriptions where almost every SKU needs a personal touch, hiring a content writer will help. Even in such cases, most of the upstream and downstream processes like product classification, specification, sourcing digital assets, content extraction, etc. can however still be automated using AI / ML based models. In most other cases, unique and relevant descriptions can be created end-to-end using machine learning algorithms. AI / ML practitioners generally use template-based descriptions with accurate attributes or apply natural language programming (NLP) to generate full descriptions or auto-complete partially written descriptions using either Markov chain or Neural net models.

Quantity and quality of training dataset

The quantity and quality of the training dataset is crucial for AI / ML models to deliver consistent results. AI / ML practitioners need to evaluate the quality and quantity of the dataset and likewise define a threshold of the model-based output. For instance, set expectations that a good quality training dataset will deliver 70%-80% accurate results which would improve over time with regular validations. Also, a significantly small dataset will be of no use to explore all possible outcomes from the model. Further, dictionaries/libraries can be built for repetitive attributes such as product specifications such as color, material, size, etc. This will help scale and improve model efficiency.

Right extraction technique for the different data sources

Generating product content involves extracting data from multiple sources like packaging labels, product images, pdfs, spec sheets, etc. It is important to identify and apply the right extraction technique for different data sources.AI and ML practitioners can use Optical Character recognition (OCR) to extract text from packaging labels, images, or pdfs, which is then processed using NLP techniques. Whereas trained neural net image processing models can be used to extract attributes from product images. Transfer learning is then applied on pre-trained models like Inception V3.

The pandemic has further accelerated the already accelerated need to offer ease in eCommerce. Winners will be those who successfully apply AI / ML to generate automated product content at scale.

By Guddi Rawat, Assistant Vice President, Ugam, a Merkle company

Leave a Reply

Your email address will not be published. Required fields are marked *