The latest addition to RS Components’ leading Digital platform, Quickfinder is an Artificial Intelligence (AI) Digital Assistant featuring Visual Product Recognition technology that allows automated product identification with the click of a button.
RS Components operates in over 30 countries and is the number one high service distributor of electronics and maintenance products across Europe and Asia Pacific. Quickfinder will initially be used internally as a knowledge base that will enable UK Customer Services Agents to provide quick resolutions to their customer queries, enhancing the experience and providing even easier product selection to customers.
Digital businesses who have access to large volumes of data increasingly invest in automation to provide slicker, personalised and faster online purchasing. Currently, customer service agents have to look in multiple data sources in order to resolve a customer’s query. Quickfinder works by using artificial intelligence (AI) technology to bring all of the data sources into one place so Customer Service teams can provide a fast and efficient service to our customers. The first application of AI technology will be in the form of a knowledge base that UK Customer Service agents can access to quickly resolve customer queries.
Visual Product Recognition is one of the key skills Quickfinder offers, alongside the likes of Search, automated Product Returns or FAQs. Customers can send a picture of a part they wish to identify and purchase and Quickfinder will return the best matches to the Customer Services agent in seconds. While the type of product and the quality of photograph are important, latest advances in Machine Learning mean they can offer an enhanced service to their customers.
“We’re observing the AI revolution happening right now and without realising, our customers and our people are already using this technology in so many aspects of their daily lives. It’s time RS Components offer similar benefits in relation to our Engineering offer.” said Cameron Ward, SVP of Innovation at RS Components.
He went on to say “We are committed to putting our customers at the heart of everything we do and Quickfinder is a great example of how we are leveraging new and emerging technology to continue to deliver a great service. Our people and their human touch is the most valuable aspect of our multi-channel Customer Experience.”
Quickfinder uses AI technology to simplify the simple and speed up the urgent, making both Employee Experience and Customer Experience more enjoyable. Starting with the initial set of skills limited to the UK market, RS expect to see the Digital Assistant’s skillset and geographical reach expand rapidly.
RS Components reached out to Hack & Craft to access our AI expertise and lean approach to deploying cutting edge technology within a legacy corporate environment.
“Hack & Craft’s aim for the project was to contribute to RS Component’s long term strategic goal to be the market leader in customer experience and to create a foundation upon which disruptive new customer interfaces could be built. We partnered our machine learning and image recognition experts with their innovation department, customer service specialists and sales team to form an agile innovation team” said Harry McCarney, Founder and Managing Director at Hack & Craft.
Musa Atlihan, a Data Scientist for Hack & Craft, explains how we were able to build the Digital Assistant using innovative technology:
"Chatbots are systems designed for extended conversations, set up to mimic the unstructured conversational or ‘chats’ characteristic of human-human interaction. There is currently a lot of interesting research being done to solve problems in the literature, with regards to the chatbots conversational tasks.
A neural conversational model, for example, is the one that uses a seq2seq model which is composed of a decoder-encoder structure based on recurrent neural networks. This kind of a model, with enough data, can output meaningful sentences after giving a previous sentence. But there are some drawbacks as the seq2seq model requires a large amount of data to create its own features and there is no guarantee for a coherent output. There are some good attempts to overcome the coherent output problem, like a persona-based neural conversation approach, but even after handling the incoherent outputs, it is still very hard to implement logic to control the conversational flows.
When building chatbots at a production level, DialogFlow becomes one of the best tools to build chatbots with contexts. It supports many languages, even for the ones where tokenization is not quite easy to apply, and gives perfect results on intent detection and named-entity recognition tasks.
However, when you are building a chatbot for complex tasks, as we have with the Digital Assistant for RS Components, you need to go beyond standard approaches. For example, in order to implement a conversational product search flow for the Digital Assistant, we needed additional scalable tools to handle the large product catalogue of RS components. RS Components have a large range of product catalogue having thousands of categories with about 6K attribute types for over 2M products. Therefore the variations of possible product search questions are endless. Using the product dataset of RS Components, we generated over 10 millions of randomly composed product search questions to train the product search intent including attribute and product name entities in DialogFlow, but unfortunately, DialogFlow intents were limited to a maximum of 2K input examples.
Within DialogFlow, it is also possible to use templates in hybrid mode (a rule-based + machine learning approach) to create input examples with composite entities by including all permutations of different attribute types and product names, however, there is also a limit for the number of entities (max of 250) in DialogFlow. Thus, there was no chance of creating an entity for each attribute type to ensure a structure-aware named-entity recognition within DialogFlow.
At this point, we decided to use another great tool named spaCy which was the best choice for us to build scalable state-of-the-art named-entity recognition models on the fly. Within the algorithms of spaCy, convolutional layer implementations with residual connections serve for a better efficiency than standard BiLSTMs. Therefore spaCy gets remarkable performances on most popular benchmarks. After training spaCy's named-entity recognition model with our huge number of product search questions, we got an accuracy of over 97% on the attribute and product name entity recognition. We therefore took advantage of spaCy by integrating it into the product search flow to make named-entity recognition more reliable.
While cutting-edge research on dialog systems is still being developed for building chatbots on a production level, implementing context-based state machines is still the best choice to have full control over the flows. However, as this project for RS has proved, building scalable solutions with the help of innovative technology, becomes an absolute necessity when a complex task with a specific dataset needs to be solved."
Click here to find out more about RS’ Digital Assistant.
Martin Peschke 1980 – 2018
To enable comments sign up for a Disqus account and enter your Disqus shortname in the Articulate node settings.