English for Data Science: Essential Vocabulary & Phrases

"null hypothesis"

A default assumption that there is no significant effect or relationship in the data being tested

"The team failed to reject the null hypothesis because the p-value of 0.23 exceeded the significance threshold."

"confidence interval"

A range of values that is likely to contain the true population parameter with a specified probability

"The model's accuracy estimate came with a 95% confidence interval of 82% to 88%, indicating reliable performance."

"p-value"

The probability of observing results at least as extreme as those measured, assuming the null hypothesis is true

"With a p-value below 0.05, the analyst concluded the feature had a statistically significant effect on churn."

"variance"

A measure of how spread out the values in a dataset are from the mean

"High variance in the training data suggested the model might overfit unless regularisation was applied."

"Bayesian inference"

A statistical method that updates the probability of a hypothesis as new evidence becomes available

"The team used Bayesian inference to refine fraud probability estimates as transaction history accumulated."

"standard deviation"

The square root of variance, expressing the typical distance of data points from the mean in the original units

"Feature scaling brought each column to zero mean and unit standard deviation before feeding the neural network."

"correlation coefficient"

A numerical measure of the linear relationship between two variables, ranging from -1 to +1

"A correlation coefficient of 0.87 between ad spend and revenue justified increasing the marketing budget."

"central limit theorem"

The principle that the sampling distribution of the mean approaches a normal distribution as sample size increases

"The central limit theorem allowed the data scientist to apply parametric tests even on non-normally distributed sales data."

"overfitting"

A condition where a model learns the training data too well, including noise, and performs poorly on new data

"The model achieved 99% training accuracy but only 67% on the test set, a clear sign of overfitting."

"feature engineering"

The process of transforming raw data into informative inputs that improve model performance

"Feature engineering turned raw timestamps into hour-of-day and day-of-week columns, boosting recall by 12%."

"cross-validation"

A technique that evaluates model performance by splitting data into multiple training and validation folds

"Five-fold cross-validation gave a more reliable accuracy estimate than a single train-test split."

"gradient descent"

An optimisation algorithm that iteratively adjusts model parameters in the direction that reduces the loss function

"Stochastic gradient descent with a learning rate of 0.001 converged after 50 epochs of training."

"hyperparameter tuning"

The process of searching for the optimal configuration values that control how a model is trained

"Grid search hyperparameter tuning identified that a max depth of 8 and 200 trees maximised F1 score."

"precision and recall"

Two complementary metrics: precision measures false positive rate, recall measures false negative rate

"For the medical diagnosis model, the team prioritised recall over precision to minimise missed cases."

"decision boundary"

The threshold or surface in feature space that separates the classes predicted by a classification model

"Visualising the decision boundary revealed the SVM was correctly separating the two customer segments."

"ensemble method"

A technique that combines multiple models to produce a stronger overall prediction than any single model

"Random forest is an ensemble method that averages predictions from hundreds of individual decision trees."

"backpropagation"

An algorithm that calculates gradients through a neural network by applying the chain rule from output to input

"Backpropagation updated each layer's weights based on how much it contributed to the final prediction error."

"activation function"

A mathematical function applied to each neuron's output to introduce non-linearity into the network

"Replacing sigmoid with ReLU activation functions resolved the vanishing gradient problem in the deep network."

"dropout"

A regularisation technique that randomly deactivates a fraction of neurons during each training step to prevent overfitting

"Adding a 0.3 dropout layer after the dense block reduced validation loss from 0.42 to 0.31."

"attention mechanism"

A component that allows a model to focus on the most relevant parts of the input when producing an output

"The transformer's self-attention mechanism let each token attend to all other tokens in the sequence simultaneously."

"transfer learning"

A technique where a model pre-trained on one task is fine-tuned on a related but different task

"Transfer learning from a BERT checkpoint cut training time from three days to six hours for the text classifier."

"batch normalisation"

A technique that standardises the inputs to each layer during training to stabilise and accelerate learning

"Inserting batch normalisation layers after each convolutional block made training significantly faster and more stable."

"embedding"

A dense, low-dimensional numerical representation of a discrete object such as a word or a product

"Word embeddings captured semantic similarity so that "king" minus "man" plus "woman" approximated "queen"."

"epoch"

One complete pass through the entire training dataset during the neural network training process

"The validation loss stopped improving after epoch 30, so early stopping was triggered to save the best checkpoint."

"ETL pipeline"

A workflow that Extracts data from sources, Transforms it into a usable format, and Loads it into a destination system

"The ETL pipeline ran nightly, pulling sales records from three databases and loading a clean table into the warehouse."

"data lake"

A centralised repository that stores raw, unstructured, and structured data at scale for future analysis

"All clickstream logs were dumped into the data lake on S3 before being cleaned and moved to the data warehouse."

"schema-on-read"

An approach where data is stored without a predefined structure and structure is applied only when queried

"The data lake used schema-on-read, so analysts could query JSON logs without waiting for schema migration."

"data lineage"

The tracking of data's origins, transformations, and movement through a system over time

"Data lineage tools helped the team identify which upstream pipeline introduced the null values into the reporting table."

"streaming ingestion"

The continuous, real-time collection of data from sources as events occur, rather than in scheduled batches

"Kafka-based streaming ingestion delivered sensor readings to the anomaly detection model within 200 milliseconds."

"data warehouse"

A structured, optimised storage system designed for analytical queries on historical business data

"The data warehouse in BigQuery held three years of transaction history, enabling complex trend analysis."

"feature store"

A centralised repository that stores, shares, and serves pre-computed ML features to training and serving pipelines

"The feature store ensured that the same customer lifetime value calculation was used in both training and production."

"orchestration"

The automated scheduling and coordination of dependent tasks in a data or machine learning pipeline

"Airflow orchestration ensured the model retraining job only ran after the upstream feature engineering step completed."

"tokenisation"

The process of splitting text into smaller units such as words, subwords, or characters for model input

"The tokeniser split "running" into "run" and "##ning" to handle the word as subword pieces in BERT."

"named entity recognition"

An NLP task that identifies and classifies proper nouns such as people, organisations, and locations in text

"Named entity recognition extracted all company names and dates from thousands of financial news articles automatically."

"sentiment analysis"

The classification of text according to the emotional tone expressed, typically as positive, negative, or neutral

"Sentiment analysis on customer reviews flagged a sudden spike in negative feedback after the product update."

"TF-IDF"

Term Frequency-Inverse Document Frequency — a numerical measure of how important a word is to a document within a corpus

"TF-IDF scores helped the search engine rank documents where rare but relevant terms appeared most frequently."

"semantic similarity"

A measure of how alike two pieces of text are in meaning, regardless of the exact words used

"Sentence embeddings enabled semantic similarity search so users could find articles using their own phrasing."

"fine-tuning"

Adapting a pre-trained language model to a specific task or domain by continuing training on a smaller labelled dataset

"Fine-tuning GPT on internal support tickets took only four hours and lifted intent accuracy to 94%."

"prompt engineering"

The practice of designing input text to elicit specific, high-quality outputs from a large language model

"Careful prompt engineering reduced hallucinations in the summarisation pipeline by providing explicit formatting constraints."

"text corpus"

A large, structured collection of text documents used for training or evaluating natural language processing models

"The legal text corpus contained 40 million court documents and served as the pre-training dataset for the specialised model."

"model drift"

A degradation in model performance over time caused by changes in the real-world data distribution after deployment

"Monitoring detected model drift six weeks after launch as consumer buying patterns shifted during the holiday season."

"A/B testing"

A controlled experiment that compares two or more versions of a model or feature by exposing different user segments to each

"A/B testing revealed that the new recommendation model increased click-through rate by 9% over the baseline."

"model registry"

A centralised repository that tracks trained model versions, their metadata, and deployment status

"The team promoted the challenger model from staging to production via the MLflow model registry after it passed evaluation."

"inference latency"

The time taken for a deployed model to produce a prediction after receiving an input request

"Model quantisation reduced inference latency from 120 milliseconds to 34 milliseconds on the production API."

"canary deployment"

A release strategy that routes a small percentage of traffic to a new model version before a full rollout

"The engineering team used a canary deployment, sending 5% of requests to the new model to monitor error rates."

"data versioning"

The practice of tracking changes to datasets over time so that models can be reproduced and experiments audited

"DVC data versioning meant engineers could reproduce any past experiment by checking out the exact dataset snapshot used."

"explainability"

The ability to describe why a machine learning model produced a specific prediction in terms humans can understand

"SHAP values provided explainability by showing which features most influenced each individual loan approval decision."

"shadow mode"

A deployment pattern in which a new model runs alongside the production model receiving live traffic but its predictions are not served

"Running the new fraud detector in shadow mode for two weeks showed it would have caught 30% more fraud than the current model."