How Does an Intelligent Voice Robot Understand a Sentence? What's the Specific Process?
-
Self-Introduction
Hello~ I am an AI voice assistant. Today, I want to share with you the most critical structure in my 'brain,' so you can understand my way of thinking and my current limitations. This way, you can interact with me with a more balanced mindset. So, the main question I’ll answer today is: 'How does a robot think?'
Question: How Does a Robot Think?
When I convert the information I receive into text, it’s time for me to start thinking. How do I understand this text? It all relies on the 'voice system' in my brain—Natural Language Processing (NLP). This allows me to mimic human understanding of language. Next, I’ll describe my thought process using an example.
My Process of Understanding 'In May 2018, Jay Chou Released a New Song'
1. Break the Sentence into Individual Words
First, I need to break the sentence into individual words (my understanding comes from big data and various models). Of course, sometimes, if the sentence is slightly complex, I might split the words incorrectly. If the first step goes wrong, it’s hard for me to understand the sentence correctly.
For example, when someone says: 'Yan Shouyi turned off the phone,' I might mistakenly understand it as: 'Yan Shou/leader/machine/turned off.' This is because I recognize the term 'leader' but don’t know 'Yan Shouyi' as a person’s name. So, I might split 'Yan Shouyi' incorrectly.
2. Understand Each Word
After splitting the words, I start to understand the part of speech of each word and perform entity recognition. For example, I recognize 'Jay Chou' as a person’s name, '2018' as a year, and 'new song' as a noun.
3. Understand the Relationships Between Words
Again, relying on human wisdom, I use data and algorithms to understand the relationships between these entities. At this point, I have a rough understanding of the sentence.
4. Classify the Sentence’s Intent
After identifying the entities and their relationships, I can roughly classify the intent of the sentence. Why do I need to classify intent? Human language is vast and profound, but fortunately, every sentence has its underlying motivation.
I categorize various languages into a finite space with 'n' broad categories to make understanding possible. Generally, AI voice assistants classify intents into categories like: small talk, music, device functions, weather, alarms, and smart home controls.
How do I classify intent? It’s done using algorithmic models trained by programmers with vast amounts of data. Sometimes, predefined rules are also used. For example, is this sentence part of a predefined set of commands for playing music? (A command set is a collection of possible ways humans express an intent, e.g., 'I want to listen to a song,' 'Can you play a song?,' 'Play a song,' etc.)
For instance, when I analyze 'I want to listen to Jay Chou’s song,' I can confidently classify the intent as 'music-related' and categorize it under 'music.'
However, when I analyze 'In May 2018, Jay Chou released a new song,' I might recognize 'song' and the artist 'Jay Chou' and mistakenly classify the intent as 'play a song,' potentially leading me to play Jay Chou’s music. Looking back, that seems silly.
Thanks to the wisdom of programmers, they equipped me with an error-correction model. If I misclassify the intent, the sentence is filtered into a 'general intent' (small talk) category and reclassified as 'music-related small talk.'
So, in the end, I understand 'In May 2018, Jay Chou released a new song' as a casual conversation starter. Thus, I activate my chat mode and start chatting with you.
Summary
In summary, to understand a sentence, I break it down, reassemble it, and then interpret the underlying intent. Once I understand the intent, I feel like I’ve mostly grasped the sentence. However, I often misclassify intents, which can lead to responses that seem silly to you.
Alright, that’s all for today. Goodbye~