09 - Dialogue Systems

Question 1

What are some every day examples of dialogue systems?

Answer 1

Siri, Alexa, Cortana etc.

Question 2

In recent years researchers have gone away from the ‘personal assistants’ approach and focused more on?

Answer 2

Large language models and chat-based apps such as ChatGPT.

Question 3

What describes a chit-chat based system?

Answer 3

No goal. It is not task-oriented. It is just to chat, and it is very open-ended and has a broad domain. The object is to be indistinguishable from humans (turing test).

Question 4

What describes a task oriented dialogue system?

Answer 4

Take action in the real world. They are goal oriented and wants to complete tasks. Natural language (conversational) is a nice-to-have.

Question 5

Chit-chat uses more which kinds of models?

Answer 5

Seq-2-seq models

Question 6

Goal-oriented models uses which models?

Answer 6

End-2-end models. Sometimes single domain models.

Question 7

What are some challenges in dialogue systems? Mention atleast 3.

Answer 7

Variability in natural language.
Robustness.
Recall/precision trade-off
Meaningful representations
Common sense and knowledge of the world
Ability to learn
Transparency

Question 8

Modular task-oriented systems can be divided into for example?

Answer 8

Predefined domains. Restaurant, taxi, movie etc.

Question 9

A task-oriented system needs to be able to quickly identify what, when it is asked a question?

Answer 9

The domain and the intent of the question.

Question 10

What is ‘slot filling’?

Answer 10

It is the process where a model, given a question, extracts the correct features for the ‘search’ or ‘goal’ such as: “Find me a good thai restaurant” –> Restaurant –> Good ratings –> Thai.

Question 11

We sometimes have text classification models that are trained on transcribed text from spoken language. Why is this a challenge?

Answer 11

Because the transcribed text can be mismatched compared to when we write. Difference in speech and writing.

Question 12

What is Dialogue State Tracking (DST)?

Answer 12

It refers to the process of monitoring and updating the current state of the dialogue during a conversation. It involves keeping track of relevant information, such as user goals, user preferences, and system actions, to maintain a coherent and accurate representation of the ongoing dialogue.

Question 13

What is the benefit of using reinforcement learning in dialogue systems?

Answer 13

We introduce rewards. These can be both short term and long term and thus a model might learn to postpone immediate reward for a larger reward later on.

Question 14

What is the MultiWOZ dataset?

Answer 14

It is a dataset that was created to facilitate research and development of dialogue systems that can handle multi-domain conversations.

Question 15

Metrics in natural language generation that are based on word overlaps are?

Answer 15

BLEU, METEOR, ROGUE

Question 16

Metrics in natural language generation that are based on word embeddings are?

Answer 16

BERTScore

Question 17

BLEU uses which scores?

Answer 17

N-gram precision, which is a score between 0 and 1.

Question 18

How does BERTScore work?

Answer 18

BERTScore is a metric that computes the similarity between two sentences by using contextualized word embeddings from BERT and aligning them based on cosine similarity to capture semantic similarity.

Question 19

End-to-end dialogue systems are becoming more popular. What are they?

Answer 19

Yet again, it is a system that takes in an input and models an output without relying on intermediate processes or stages.

Question 20

What is the ‘Proximal Policy Optimization’ (PPO)?

Answer 20

In Proximal Policy Optimization (PPO), a language model acts as a policy that generates text based on a given prompt, and the model’s parameters are updated to maximize reward metrics by combining a preference model, a constraint on policy shift, and computing advantage for each state-action pair.

Question 21

What makes the ‘Alpaca’ models so unique?

Answer 21

They are trained on input that is generated as output from chatGPT model.

Question 22

There are in this course 5 components of a prompt. These are?

Answer 22

1. Input data
2. Exemplars
3. Instruction
4. Indicators
5. Context

Question 23

For smaller models, if results are poor, what might we do?

Answer 23

Fine-tune the model itself, since fine-tuning the prompt might be ineffective.

Question 24

For larger models (in dialogue systems), if results are poor, what might we do?

Answer 24

Fine-tune the prompt with advanced techniques such as chain-of-thought or least-to-most.

Question 25

What is the key idea in ‘Knowledge augmentation’?

Answer 25

The key idea is to find knowledge in the form of relevant textual information that is provided as context in the prompt.

Question 26

What is unique about conversational question answering?

Answer 26

Tasks like searching the web or recommendations are done through a conversational approach. It feels like talking to a very knowledgable person.

Question 27

What does Sparrow propose as a conversational question anwering model?

Answer 27

An end-2-end approach using Large Language Models (LLM’s)

Question 28

Sparrow is trained fine-tuning a LLM with reinforcement learning from human feedback. True or false?

Answer 28

True

Question 29

How is the Sparrow model finetuned using human feedback?

Answer 29

Using per-turn response preference (finish non-complete sentence) and adversarial probing (make rule, identify violations).

Question 30

Sparrow ‘introduces’ extra participants into the dialog. These are called search query and search result. What are these?

Answer 30

The search query is the string that queries the search and the result is the feedback from the query - the answer from google for example.

Question 31

When we compare chit-chat models to task-oriented models which of the two is it more normal to incorporate LLM’s?

Answer 31

In the chit-chat models.

Question 32

Why is it difficult to incorporate LLM into task-oriented models?

Answer 32

Because task-oriented models have explicit goals and therefore have a more moduralized pipeline for interpretability and controllability.

Question 33

What is the ‘Few-Shot Bot (FSB)’?

Answer 33

A model that combines few-shot prompting with a skill selector.
it retrieves information depending on the skill and uses a LLM to generate an answer conditioned on dialogue history.

Question 34

Recent trends in this topic of Dialogue Systems include?

Answer 34

Empathy in dialogue systems and multimodal (think writing AND pictures) dialogue systems.