15786136
Description
Flashcards by August Edström, updated more than 1 year ago
|
Created by August Edström
about 6 years ago
|
|
| Question | Answer |
| What is the main difference between a chatbot and a dialogue system? | There is no goal-orientation in a chatbot, whereas in a dialogue system there is. So the difference lies in the result of the different conversational agents. |
| Briefly explain how the Eliza chatbot works | A chatbot. No specific goal/task. Looks for keywords which then have associated pattern/transform rules. Given a sentence(pattern), a response to user based on pattern(transform). Standard answers and a memory trick |
| What is the difference between an IR-based chatbot and a seq2seq chatbot? | Seq2seq only maps words vs IR-based maps sentences. IR will not say a sentence that is not in the training corpus |
| Briefly explain 2 ways in which a seq2seq model for a chatbot can be improved | Use previous turns as context Beam search: look not just at the next best words (greedy) but at the next best sequences Use reinforcement learning to make conversation as a whole more natural |
| What are the pro’s and cons of chatbots? | Pros: Fun, applicaitions to counseling, good for narrow, scriptable applications Cons: They dont really understand, rule-based chatbots are expensive(time-wise or computing-wise) and breaks easily, IR-based can only mirror training data |
| What are the components of a typical frame-based dialogue system? | Speech recognition, NLU, Dialogue manager, Task manager, NLG, Text-to-Speech Synthesis |
| What are the disadvantages of a simple finite state dialogue manager? | Completely system initiative. I.e. system decided what's important. Limited use. Real dialogues involve give and take |
| What are the advantages of dialogue system with system initiative? | Easy to build, user know what's expected, system know what user will say next, OK for simple tasks |
| What is a frame and how can it be used to guide a dialogue? Provide a concrete example. | A frame has a set of slots to be filled with values of certain types. Each is associated with a question. User can answer multiple questions at with one answer. For example, when booking a flight: If the user gives information that can fill multiple slots then the system dont have to ask again about those slots. |
| Briefly describe how machine learning can be used to learn domain and intent determination, slot presence and slot filling | Domain and intent: used to map words to semantic frame-fillers() and use Slot presence and slot filler: 1-on-N classifier(naive Bayes, logistic regression, neural network, etc.)) |
| Name 3 possible components of a dialogue state | Slots, wheter slots have been filled/confirmed, user's dialogue acts and how they want to interact with the system |
| What is a dialogue act? Provide 3 examples. | An act with a structure related specifically to its dialogue funtion(a "conversational move"). For example: Hello(), Inform(), Bye() |
| Can we determine user dialogue acts using just the surface syntactic form? Why (not)? | No, because there may exist ambiguity about an act, that is, what looks like a question may be a request |
| How can we evaluate a dialogue system using slot error rate? | By looking at the error rate of the information in slots filled by a sentence |
| What two chatbot architectures are there? | Rule-based & corpus-based |
Want to create your own Flashcards for free with GoConqr? Learn more.