Case Studies

Improving Chatbot Conversationality for Linde


About Linde

  • $27B annual revenue (2020)
  • 100 countries
  • 194 offices
  • 74,000+ employees
  • World’s largest industrial gas and engineering company with 100+ years of history

Problem: Voice chatbot failing to understand user queries

Linde leveraged Super.Classify to improve its employee experience by building a smarter voice-based chatbot, demonstrating how artificial intelligence (AI) can empower people to communicate with machines using natural language and accomplish their goals in as few steps as possible.

The company operates globally in the industrial gas and engineering sectors, serving a variety of end markets including chemicals and refining, food and beverage, electronics, healthcare, manufacturing, and primary metals. While building a chatbot to act as an internal knowledge base for its employees, Linde had difficulty training the bot to recognize user intent and named entities in queries. Accurate intent recognition is the crux of an effective chatbot—without it, users are served irrelevant information. Working with super.AI, Linde solved this problem by using artificial intelligence to improve its chatbot’s understanding of natural language.

We were looking for a supplier with named-entity recognition (NER) expertise that could provide us with high quality training data to feed into our existing ML service we were using to build a chatbot. We were very happy with the throughput and speed we received from super.AI

Solution: Improve natural language understanding with AI

Using Super.Classify, super.AI’s no-code solution for data classification, Linde was able to quickly create an accurate training dataset and build a better chatbot. Using a combination of AI-powered data classification and human verifiers, super.AI created a data program capable of accurate natural language processing (NLP) and named entity recognition (NER). 

To begin, Linde provided a list of entities, utterances, and ground truth. Super.AI then classified the utterances and provided a compatible output. Our approach involved several steps:

  1. Initial processing 
  2. Variable feedback processing
  3. Text augmentation 
  4. Query augmentation with synonyms
  5. Data augmentation

For an initial proof of concept (POC), a snippet of Linde's internal database that included 50+ columns and 20+ rows was provided. The rows in the database are individual projects and the columns are the properties of interest. To generate input for Linde's custom data program, two types of queries were created. One that targets either one or several individual columns for one specific row of the database (e.g., What is the number of users for project 'X?') or queries that filter rows by column conditions (e.g., How many projects have more than 10 users?)

The data program performs text augmentation on these inputs. Given a query, the data program produces many similar queries with the same intent but different wording. To accomplish this, the data program uses three different strategies:

  • Swapping single words with synonyms from a thesaurus
  • Back translation or reverse translation
  • Involving human workers to both generate phrase variations and verify them against the original input for intent similarity

Linde now has an internal knowledge base that quickly responds to spoken user queries posed in natural language.

Project Highlights

  • Produced accurate training data in 6 weeks from kickoff to final delivery
  • Combined AI and human workers to achieve high data accuracy


Get a customized demo with your documents

Book a free consultation with our experts.

Related Case Studies

Button Text