Overview

As per the course website, "This week’s lecturecast focuses on one of the most rapidly growing technologies in intelligent agents, which is Natural Language Processing (or NLP). The lecturecast covers some of the background to NLP and why it is hard for machines to understand the way that we talk before diving into some examples of methods and technologies that can be deployed."

My Reflection

Overall Reflection

This first week of the second half of the module moves on to the topic of Natural Language Processing (NLP). I hope that this would be a transition towards approaching the modern-day agents, after spending the first six weeks mainly focusing on symbolic agents, satisfying my eagerness and feeling of the inevitable coverage of language models-based agents in a true MSc program module about Intelligeng Agents today. But frankly, after all of my experience with this program, I doubt there will be such a transition. My doubt is doubled by my disapointment from the fact that this unit's materials did not even mention large lanuage models (LLMs) and all the advancements that they have been witnessing lately. Therefore, this unit -and module overall- would have been perfect to study in 2016. Unfortunately, I am studying it in 2026.

The unit included a lecturecast, three papers to read and a requirement for a summary post for the second collaborative learning discussion, started in Unit 5. The lecturecast focused on a few historical approaches to make machines capable of dealing with natural language, why that is a difficult task. It went through the difference between syntax, semantics and pragmatics of any language, and mentioned some characteristics that make natural language tricky to be 'understood' by machines. These characteristics include ambiguity, vagueness, polysemy (a word can have multiple word senses depending on the context) and how sentences can be grammatically (or syntactically correct) but still have no meaning (no semantics or pragmatics). I found Noam Chomsky's example of "colourless green ideas sheep furiously" interesting and I intend to look more into it in the future. The lecturecast also touched on different approaches of machine natural language processing approaches, like Word2Vec, Google News Corpus and maths of language, Hearst Patterns. It also introduced the concept of hyponymy, for which there have been several approaches to automatically detect, and which is also covered by one of the papers in the readings list.

Readings Reflection

The first reading paper, by Mikolov et al. (2013), which was a follow-up on a previous paper introducing the skip-gram model, the model that introduced the possibility of predicting the neighbours of a word. This paper introduced three main enahncements on the model:

1. Negative Sampling (NEG): to deal with noise and irrelevant context that is cheaper and faster than the originally proposed hierarchical softmax method.
2. Subsampling of Frequent Words: to deal with frequent words, like stop words, that are the most frequent and yet provide the least information. This method makes sure to not deal with those words equally to the actual words with meanings, and to drop them from the computation. As far as I understand, before this enhancement, skip-gram model treated frequent words equally with actual words.
3. Phrase Vectors: dealing with phrases rather than single words only, enabling the detection of phrases, like "Boston Globe" as a newspaper instead of treating it as a natural compositional combination of "Boston" and "Globe".

These enhancements in the skip-gram model was the foundation upon which the word2vec package usable at scale, extending as one of the roots to the evolution of NLP leading to Generative Pre-trained Transformers (GPT) and their continuous development since 2017.

The second paper was the original paper in which Marti Hearst introduced his Hearst Patterns to detect hyponyms. It was interesting for me to see how he was able to formulate logical rules to detect and extract hyponyms in natural language, and I was able to see this must have required a deep understanding of linguitics and semantics in addition to computer science.

The third paper was a more recent one by Aqab and Tariq (2020) that introduced the use of Convolutional Neural Networks (CNNs) in a workflow of pre-processing and post-display as an enhanced technique for English handwriting recognition. Before the introduction of this approach by this paper, the previous approaches included implementing Hidden Markov Model (HMM) statistical model, and other machine learning models like Support Vector Machines (SVM), as well as a few attempts with Artitifical Neural Networks (ANNs). It was nice to revisit CNNs, deep learning and machine learning in general for the first time in the program for the first time since the Machine Learning Module. It was also worthy to notice for me that the paper was based on an empirical study, comprising university students and professors, which I did not see very frequently in similar computer science studies.

Aqab, S. and Tariq, M.U. (2020) ‘Handwriting Recognition using Artificial Intelligence Neural Network and Image Processing’, International Journal of Advanced Computer Science and Applications, 11(7). Available at: https://doi.org/10.14569/IJACSA.2020.0110719.

Hearst, M. (1992) ‘Automatic Acquisition of Hyponyms from Large Text Corpora’, in COLING. The 14th International Conference on Computational Linguistics. Available at: https://aclanthology.org/C92-2082.pdf (Accessed: 13 March 2026).

Mikolov, T. et al. (2013) ‘Distributed Representations of Words and Phrases and Their Compositionality’, arXiv (Cornell University) [Preprint]. Available at: https://doi.org/10.48550/arxiv.1310.4546.