Deliverable 1. Interactive NLP-based AI system

You should build a dialogue-based system which allows a user to complete a transaction (see below) with an AI, commonly known as a chatbot, or a conversational AI. The work will be based on what you have learned, using Python and a restricted set of allowed libraries1 . A lot of Python libraries can automate much of the process, so it is of the utmost importance that you do not use them. Any part of the assessment that is performed by a library will not be counted towards your grade.

The two key objectives for this coursework are:

.    Natural  language  processing  architecture:  the  system  needs  to  include  the  architectural features that show what you have learned about processing text in lectures and labs.

.    Conversational design: the system should implement the conversational design principles as covered in lectures to ensure a usable and satisfactory user experience of your chatbot.

The system’s main functionality should be focused on supporting the user in making a transaction, i.e., it should support users in completing a task, such as booking and ordering in a specific domain of your choice. Some examples (not an exhaustive list) of acceptable tasks include:

.    Travel booking

.     Restaurant booking

.    Cinema/theatre/gig ticket booking

.    Groceries and home delivery ordering

.    Cooked food home delivery ordering

.    Cook-along recipe provision

.     Uni timetable management

.    Calendar management

.     Kitchen inventory management

.     Playlist management

.    Your turn, be creative! If you are in doubt whether your idea satisfies the requirements, ask us in the lab.

1.1   Architecture

This  section  contains  a   non-exhaustive  list  of  common  tasks   in  interactive  NLP-based  systems (chatbots) that you are expected to implement. They are not all equal in complexity or importance, e.g., intent matching is necessary for any multi-usage chatbot while small talk is not.

Expected features:

1.  Intent matching

Intent matching is a necessary part of any system which aims to do more than one task. It consists in taking in a user input and predicting the intent of the user, i.e., which function of the system the user is intending to use. Such functions could be storing/showing the user’s name, outputting the current date and time, starting a transaction dialogue, or anything else described in this document. Intent detection can be done at different degrees of complexity, from simple text matching to similarity- based and machine learning-based approaches.

2.  Identity management

Identity management refers to being able to acquire, memorise, and use the name of the person currently using the system. It consists in functions such as detecting name statements (e.g., “my name is X”, “call me X”, etc.) and storing the name to be used either in basic statements addressing the user, or in explicit name output functions (e.g., “what is my name?”).

3.  Transactions

A transactional dialogue system is a tool that allows for an artificial system to dialogue with a human operator in order to achieve a specific set of goals, which are transactional in nature, such as booking a train/plane ticket, a table in a restaurant, ordering an item from a store, etc.

Here is an example of a turn-by-turn dialogue between an operator and a travel booking system, to illustrate the range of tasks that such system would need to accomplish to properly function.

Turn    Author    Text

4.  Information retrieval & question answering

A conversational information retrieval system is an information retrieval tool that interprets a query from a dialogue with a user and returns a set of one or more results which are deemed to be relevant, such as direct answers, webpages, books, recipes, songs, movies, etc. It requires having some form of database of documents and an indexing system. A question answering system can be seen as a special case of information retrieval where the documents are question-answer pairs, and only the answer is retrieved.

5.  Small talk

A small talk function refers to the ability to have short discussions about shallow topics, such as “hello”, “how are you”, “how’s the weather”, etc. There is no expectation of long conversations due to them  being one-off topics, but they can either be single turn (question, answer) or multi-turn (question,  answer, and one or more follow-ups).