eva Automated Learning
For questions less frequently asked but still relevant to the virtual agent's context, there is an application in the Dialog Manager that dispenses intent training: Automated Learning.
Imagine you're working on a virtual agent project for a Bank and you need to add a lot of institutional information, such as the history of the bank, CEO's profile, the board of directors, legal documents, etc... Although these information are important, customers rarely ask for it.
In those cases, you can use Automated Learning (AL), a question-answering software that interpret users’ questions and answers them based on documents uploaded beforehand.
The eva AL, cognitive engine developed by NTT DATA, helps you save time and expands your knowledge base, independently of the NLP you're using.
How does Automated Learning work
Finds an answer to the user's question in one or multiple documents, without previously training any data, like it's needed for the intents engine
Has a FAQs functionality for documents by registering pairs of questions and answers. Those answers are brief and accurate
Doesn't require specific training to work with a text knowledge base. Automated Learning is multi-language, and it has been tested in Spanish, English and Portuguese
Generates an output (likelihood) indicating the confidence of the model response
Has the ability to identify if there is no answer to a question, given the context of the document
Important: Frequently accessed subjects and transactional subjects need to be accessed by the Intent engine
Although Automated Learning can read and extract answers from any text, some best practices make it more assertive. Read them below.
a) Longer paragraphs work better than short ones
Automated Learning works better with context.
The longer the material is, the better the inference is. Paragraphs with four sentences work better than paragraphs with two sentences. If you have two small paragraphs, it's better to put them together and make it longer, as a four-sentence paragraph. For example:
John is a nice man. John lives in New York. John likes zucchini.
John is a nice man. He lives in New York. John likes zucchini.
b) Avoid bulleted lists
Bullet lists are the extreme example of the lack of context that small paragraphs have. Try to make a paragraph of those lists. For example:
- Gazpacho
- Tomato
- Onion
- Garlic
- Stale bread
- Salt, olive oil and vinegar
To make gazpacho, you will need tomato, onion, garlic, stale bread, salt, olive oil and vinegar.
If you have a large list, try to split it. For example, if you have a list with 30 items, you can split it into three lists in three different paragraphs, like this:
List A – Items 1 to 10
List B – Items 11 to 20
List C – Items 21 to 30
c) Avoid tables and forms
Automated Learning does not read well tables and forms. It is better to rewrite the content as paragraphs.
d) Remove special characters, emojis, and images. Non-ASCII characters tend to turn to strings of characters that are not readable by Automated Learning. Although some special characters are common (like the dollar sign $), some are more obscure and can hamper the NLP inference, like the character ♠. Emojis count as special characters and should be removed as well. Images also might interfere with eva NLP AL functioning.
e) Remove headers, footers, summaries, references and any other elements that are not relevant to the document.
Headers, footers, summaries, and elements alike add unnecessary information that might make it harder for the cognitive engine to deliver precise answers.
f) Pay special attention to punctuation
Errors in punctuation hamper understanding much more for a cognitive engine than for a human reader. Punctuation matters a lot. “I’m sorry; I love you” is an apology. “I’m sorry I love you” is cruel.
g) Avoid passive voice
“Gazpacho is an example of cold tomato soup” works better than “An example of cold tomato soup is gazpacho”.
h) Avoid subordinate clauses
If you have a sentence with subordinate clauses, try to dismember it in smaller sentences. “Gazpacho is a cold tomato soup from Spain that, although characterized by the tomato, didn’t took tomato until the 19th century” is a long sentence with lots of subordinate clauses. “Gazpacho is a cold tomato soup. It is from Spain. Although characterized by the tomato content, it didn’t take tomato until the 19th century” works better.
i) Pay attention to the voice and tone of your documents
As the answer inference is automatic, if you extend your virtual agent voice tone to its documents, you will guarantee a consistent virtual agent personality.
Some tips when uploading FAQs (Frequently Asked Questions)
Before uploading a FAQ on AL, make sure that you have removed the questions and upload only the answers. Automated Learning works on FAQs better if the questions are absent.
When writing a FAQ, think of your text as a question-answers pairs. First, write them and then remove the questions.
Write answers to short questions. Not only most doubts can be expressed with short questions, but AL works better when the questions are short.
Using Automated Learning
After you have adapted your documents following those guidelines and converted them to .txt, you can upload them to Automated Learning. Clicking on the Automated Learning tab in the Cockpit will lead you to the Automated Learning first page.
Important: Enabling this feature may result in additional costs for each new request.
Follow these steps:
Enable Automated Learning, as it comes disabled by default.
Upload documents (TXT format).
To import a document, click on “Import document”. After the first document is imported, you will be taken to the documents’ repository.
There, you can check the documents that are part of a virtual agent knowledge base, how many and which questions were attached to a particular document, delete, and disable or enable a document.
Click on the eye icon to view the content of the document. You can sort the items of the repository by its name, amount of questions attached to it, and last added.
You can enable or disable a document.
Important: When you disable a document, all questions attached to it are also disabled. When you delete a document, all questions attached to it are also deleted.
Questions
Although Automated Learning is able to find answers without the need to create questions, you can use this feature to edit or customize the answers for each channel, for example, and also to improve the inference.
Click on “Create question”, give it a name, attach a question and its variants (other ways the user may ask for that specific subject) to the right answer.
Just like the cells in the Dialog Manager, you can use tags to label the questions too.
Example:
In a hospital FAQ document, you want to attach the question “Are hospitalized patients entitled to a full-time companion?”.
Some users may ask this question in a different way like “Do inpatients have the right to a full-time companion?” or even “Can I accompany a hospitalized patient?”. There are many variants to a question. Try to add different variants to improve your inference.
After the answer is inferred, click "Next" to move on to the next step where you can curate the answer, just like you would do in a regular answer cell, but with the inferred answer in the text box.
You can copy any part of the text and paste it to the text box, or rewrite the answer as you wish. As in a regular answer, you can add buttons and/or technical text.
Example:
For the question “Are hospitalized patients entitled to a full-time companion?”, the Automated Learning found the answer "The right to a full-time companion is guaranteed, in cases of hospitalization, for the elderly and/or disabled." You can edit it so it becomes "Elderly and/or disabled people have the right to a full-time companion" and add more details.
If you want to link the answer to other text, you can change the document that the answer belongs to by clicking on “change document” above the text box. You'll see A modal with the documents uploaded so you can select a new document.
Important: every question must be attached to a document.
After you click “save”, the question-answer pair will be stored in the Questions repository.
In this repository, you can sort the questions by its name, the amount of variants, and filter them by documents.
Last updated
Was this helpful?