AI Knowledge Engineer: a key role in any AI project

[???????????? ?????????????? ???????????????-?Article by?Chiara Martino , AI Conversation Designer & Knowledge Engineer at Assist Digital]

AI Knowledge Engineers teach AI products to understand user inputs and to provide outputs accordingly.?

Their responsibilities include working with data and building smart architectures, so that the AI can simulate human decision-making.

What does a Knowledge Engineer do

Depending on the type of project, some of a Knowledge Engineer’s main tasks are:

• Data sets creation: collecting and cleaning unstructured data to build an unbiased and representative dataset.
• Data annotation: categorization and labelling of data in various formats so that machines can use it.
• Algorithm training and test: using that data to train, validate and test algorithms.
• Flows implementation: transforming the designs into rules used to teach the system what to do in different scenarios.
• Analysis and fine-tuning: analyse outputs, in order to adjust and re-train the model when the AI doesn’t classify new data correctly.

What does all this working with data and rules mean? Let’s take a look at a concrete scenario.

Knowledge Engineers in Conversational AI Projects

In the case of a Conversational AI project, so one in which the goal is to create a chatbot or a voice assistant, Knowledge Engineers will train the Conversational Interface to understand what users say or type and to answer them.

How a Conversational Interface understands users

Let’s start from the “understanding” part of the process.?

Clearly, a Conversational Interface does not understand in the same way humans do.?

Understanding, in this context, means categorizing a user utterance into the right category of meaning and extracting useful information from it.?

To let the AI do so, Knowledge Engineers transform the unstructured human language in something that an AI Assistant can process:

• First, they will collect a large number of user utterances, the different expressions in which users might express a need
• Then, they will clean these unstructured utterances from misspellings, grammar mistakes and unnecessary details, to transform them into training phrases. These will constitute the data sets that will train the NLU (Natural Language Understanding) model.
• After that, each training phrase will be assigned to an intent, which is a set of phrases that express the same user intention for a specific conversational step.
• Moreover, if needed, each phrase might be tagged with entities, sets of words or group of words which share a semantic relationship (synonyms, hyponyms, etc…).

At this point, the actual training takes place. The NLU system will use those labelled training phrases to learn to recognize also the meaning of new sentences, that it has never seen before: it will learn to associate them to the intent with the most similar training phrases.?

Let’s see an example.

A Conversational Interface for an e-commerce might include a Use Case to handle the loss of credentials to log into the private account.?

In this case, the intent could be called “credential loss” and it could be trained with sentences similar to these: “I can’t find my credentials”, “I forgot my username”, “I lost my password”.

Words like “credentials”, “username” and “password” might be tagged as different values of the same entity “credential type”, as they are semantically related in this context.

Finally, after the model has been trained, it should be able to classify with the intent “credential loss” also similar but not identical utterances, such as “I can’t find my password, so I cannot log in”.

How a Conversational Interface answers to the user

After the understanding step is all set, Knowledge Engineers need to deal with the answering part of the process, which basically means setting the rules to connect each intent to the corresponding response.?

Today, most Virtual Assistants used in business scenarios do not use NLG (automatic Natural Language Generation), as it is still considered to be too unpredictable. On the contrary, dialogues are written by Conversation Designers, who design both the logic behind a conversation and the script itself.

Conversation Designers usually hand conversational flows to Knowledge Engineer, who are called to implement them.?

In doing so, they’re required to call APIs to get data about users, products, services, use and verify variables and conditions, to activate one conversational step or another and make the bot perform operations (e.g. send an e-mail, unsubscribe from a newsletter, etc)

Testing, Deployment and Fine-Tuning

An accurate testing phase is always required, both for the NLU part and for the implementation part, to check whether new sentence are correctly classified and whether the flows implemented work correctly and respect the design.?

In this step, Knowledge Engineers might be helped by QA testers.

Finally, when the tests are over and both intents and conversational flows have been deployed, Knowledge Engineers analyse real user-bot interactions, in order to constantly improve the system and find new use cases that need to be created.

Knowledge Engineers’ skills and backgrounds

While this role is blooming with Conversational AI, Knowledge Engineers are not only required in the creation of Conversational Interfaces, but are key in various types of AI projects, such as machine translation, automatic e-mail classification, semantic search engines, sentiment analysis, and so on.

A Knowledge Engineer is a hybrid professional, that often has a background in Maths, Statistics, Data Management or Computational Linguistics.

They need to have logical, technical skills, as this profession involves working a lot with data and rule-setting, that enable AI to connect the dots. However, they do not necessarily need to be developers, as they can use low-code and no-code platforms.

But they also need to have linguistic skills to be able to set the right semantic rules and to collect data that are well-balanced and representative.

Their mixed skill set and their ability to speak both tech and human make them vital in any AI project, as they often work as a bridge between designers and developers.