It’s common for the intent used on welcome messages or beginning of a conversation with a bot to start with “how can I help?”. This is called a first state or session start. It’s best to include only phrases that actually answer this question.
Although unlikely, there are some cases where sentences do not match a first-state intent. Here are the main examples of these sentences (potential answers or incomplete variants):
- Short sentences copied from detailed intents found in further states of the bot, e.g. To the travel policy.
- Sentences without an utterance object, which have been replaced by a pronoun, e.g. I just unlocked it .
- Sentences that start with affirmations and negations, e.g. Yes, I would like to buy a policy.
- Other sentences that look like incomplete or truncated statements, e.g. I sell online. I buy and sell francs there.
Keeping these types of sentences within the intents of the first state lowers the effectiveness of the algorithm because:
- The structures are similar to sentences for other intents found elsewhere in the bot,
- They are dissimilar to the rest of the first state sentences (in terms of vocabulary and length), so they can be confusing to the algorithm,
- They do not indicate an object, which is sometimes the only element that allows to distinguish similar sentences with different intents and are less informative than other sentences in a given intent (what is meant by "it" within the phrase), and therefore constitute a classification problem also for a human.
Updated 3 months ago