Plain-language explanation.
Computational linguistics is the use of computers and mathematical methods to analyse and generate human language. It underpins technologies like Google Translate, Siri, and ChatGPT. It combines linguistic theory with machine learning and computer science to tackle problems like automatic translation, speech recognition, and information extraction.
Core concepts and standard treatment.
Computational linguistics spans: formal language theory (Chomsky hierarchy — regular, context-free, context-sensitive, recursively enumerable languages), corpus linguistics (statistical analysis of large text corpora — collocations, frequency distributions, semantic prosody), machine translation (rule-based, statistical, neural — transformer-based NMT), parsing (constituency and dependency parsing algorithms), information extraction, and question answering.
Deeper theory, debates and edge cases.
The statistical NLP revolution positioned probabilistic models over linguistic rule systems. Distributed representations (word2vec, GloVe) replaced sparse symbolic representations. The transformer architecture (Vaswani et al. 2017: attention is all you need) enabled BERT (bidirectional encoder) and GPT (autoregressive decoder) — pre-trained language models achieving state-of-the-art on most NLP benchmarks. In-context learning, chain-of-thought prompting, and instruction tuning are recent paradigm shifts.
How it is applied in practice.
Evaluation challenges in NLP: benchmark saturation motivates harder benchmarks like BIG-Bench, MMLU, HellaSwag; dataset artifacts and annotation biases inflate measured performance; and the gap between benchmark performance and real-world deployment quality is large. Multilingual NLP: most large language models are dominated by English training data — multilingual models and language-specific models address but do not fully close performance gaps for low-resource languages. Linguistic annotation (treebanks — Universal Dependencies; semantic role labelling — PropBank) is the foundational data infrastructure for supervised NLP, requiring costly expert annotation.