The following figure shows the general structure of a dialogue system (DS).

Automatic Speech Recognition (ASR) converts the user speech to the text. The text is then analyzed and understood in Natural Language Understanding (NLU). Dialogue manager keeps track of the input and the corresponding responses. Natural Language Generation (NLG) generates the appropriate responses, which are converted to the speech by text to speech. The goal of the thesis is to implement NLU for an argumentative DS. The following figure shows the NLU structure.

The NLU design is influenced by the way humans understand an utterance in a discussion. In a discussion, we argue based on our knowledge base about the topic. Richer the information we have, better is the quality of the argument. Moreover, the responses also depend on the intention of the opponent. For example, we respond differently for opponents trying to concede, argue, or retract.

The design consist of an utterance mapper, a database and an intent classifier. The database contains arguments about a topic. The utterance mapper finds the sentence closest to the user utterance in the Database. Intent classifier detects the move played by the user in an argument game. A user can have four types moves or the intents, argue, why(asking for more information about the previous statement ), retract and concede. Both these outputs are needed to synthesize the response.

The utterance mapper is implemented by various models like the ones which exploit the linguistic properties of the English language such as, WordNet, models which rely on the vector representation of words like GloVe, BERT. The performance of these models are measured with different similarity measurement techniques. The similarity measurement techniques include various distance measurements and certain aspects of IBM project debater.

We use the sample Wikipedia dataset released by IBM project debater to train GloVe. BERT trained model is downloaded from the Google release. WordNet model is downloaded from the NLTK package.

For the performance measurement of GloVe, BERT, WordNet and similarity measurement techniques of IBM project debater, we convert this dataset into a dataset of sentence triplets. A sentence triplet is a set of three sentences where two sentences are semantically closer than the third. A successfully trained model should identify the relation between the triplets.

Intent classification is implemented by the chat bot designs such as RASA.

The database is constructed according to [3]. It contains a sample debate from the database of idebate.org. Idebate.org is operated by the International Debate Education Association (IDEA), a global Network of organisations who educate people in debate.

The following packages are needed to run the scripts.

• Tensorflow 1.13.0

• Numpy

• NLTK

• BERT server and BERT client

• Progressbar

It is recommended to install these packages inside a virtual environment. Create a new virtual environment by choosing a Python3 interpreter and making a ./venv directory to hold it:

Activate the virtual environment

Install all the necessary packages.

The scripts depend on /Training_Data/dataset.csv from IBM project debater for training and validating the models. This data set contains 692 Wikipedia pages of May 2017 where, article title, section title and sentences are arranged in a CSV format. This data set is transformed into triplet sentences and triplet titles according to [2] (Can be found in /Training_Data/triplets). Following table summarizes the triplet counts generated From dataset.csv.