If a customer sends: Hey, what’s up? Do you guys offer a free trial? and your bot responds: Sorry, I don’t understand, there is a big problem with multiple intents you need to solve. Most of the chatbots and virtual assistants are unable to understand users when two different requests are included in just one sentence. How difficult is then to make your Artificial Intelligence handle more than one intent? Easy-breezy!
Most chatbot systems are built on the basis of intent and entity detection. This concept implies identifying both the intent of a request and its related entities. If a bot receives the following request: “I want info on iphone X ”, it would detect ‘want info’ as the intent, and ‘iphone X’ as an entity. For such simple requests, most chatbot frameworks will work correctly.
The problem arises when a user says: “I want info on iphone X and delivery to NY”. When customers come with more complex queries, the understanding skills of those conversational agents are usually underperforming. Thus, instead of only focusing on the appearance of a bot, it is more recommendable to work on its Natural Language Understanding abilities, which needless to say is the core element of any conversation.
All at once: how multiple intents are used in real life
In linguistics, coordination is defined as the use of coordinating conjunctions, conjunctive adverbs, or punctuation symbols to combine short independent clauses to form a whole sentence. This is a common feature dealt with by NLP techniques when performing bot training. Nevertheless, it turns out to be not that easy to parse, specifically because a lack of tools for syntactic analysis. Let’s take a look at real-world examples that can clearly illustrate that gap:
As mentioned above, users often order different items or ask about more than one product or service in one single request. In such cases, if this multiple-intent issue is not properly solved, revenue rates may be consequently reduced.
- Home automation
Using a basic common example of a home automation intent such as the management of the lights in a room can help us pinpoint the challenge. If you ask your home assistant in a single query to turn on the lights in the living room but to also turn off those in the kitchen, it will get confused and be not able to accomplish both tasks at once.
TechCrunch, the world-famous publisher of tech industry news, and its Messenger bot were launched to deliver news and stories on demand. Initially, the TechCrunch bot was just able to react to basic requests, such as asking for the main menu, managing subscriptions or giving feedback. In our post How we made TechCrunch bot more conversational you can discover how natural language skills were greatly improved by deploying Bitext NLP middleware services.
The TechCrunch bot became smarter and more conversational leveraging our innovative Query Simplification service, which fundamentally consists of rewriting user queries to make them simpler and easier for a bot to understand. A complex request such as “Can I get news about wearables and smartphones, please?” will be rewritten into two different intents ‘get + news + wearables’ and ‘get + news + smartphones’. Thus, Bitext NLP middleware acts as a convenient NLP layer to boost bots’ performance.
Bitext technology also helps conversational agents deal with other complex linguistic phenomena as negation ("I want a deluxe pizza without mushrooms") or conditional structures ("If it is included in the offer, add a beer") not needing extensive intent detection rules or hundreds of examples for training. These services are based, in turn, on a syntactic component or parser, available in more than 20 languages, and easy to integrate into any chatbot framework or search engine.