How Word Analogies Actually Work
Word analogies are a way to represent relationships between words by mapping them into vector space. The classic example is king minus man plus woman equals queen. That's not magic, it's just linear algebra on word embeddings. When you train a model on enough text, certain semantic relationships emerge as consistent directional vectors. Subtracting "man" from "king" isolates the "royalty" component while removing the "male" component. Adding "woman" back in shifts the gender dimension, landing you near "queen." I spent years working with NLP models before I really understood why some analogies work reliably and others fall apart completely. The surface-level explanation is straightforward, but the practical reality is messy. A lot of people treat word analogies like a party trick for AI. They're not. They're a diagnostic tool for understanding how well your embedding space captures relational structure.
100 Examples Of Word Analogy
Here is a comprehensive list covering the main categories. These are the kinds of analogies that tend to appear in standard benchmarks like the Word Relationship Completion dataset and Google's original word2vec paper tests. Occupation and Role: Doctor is to hospital as teacher is to school. Lawyer is to court as judge is to courtroom. Pilot is to airplane as captain is to ship. Writer is to book as composer is to symphony. Chef is to restaurant as artist is to gallery. Firefighter is to fire as paramedic is to ambulance. Architect is to blueprint as engineer is to schematic. Mechanic is to engine as electrician is to wiring. Surgeon is to scalpel as dentist is to drill. Accountant is to ledger as auditor is to audit report.
Family and Relationships: Mother is to daughter as father is to son. Uncle is to nephew as aunt is to niece. Grandfather is to grandson as grandmother is to granddaughter. Brother is to sister as husband is to wife. Son is to father as daughter is to mother. Cousin is to cousin as sibling is to sibling. In-laws are to family as step-parents are to blended family. Nephew is to uncle as niece is to aunt. Twin is to twin as sibling is to sibling. Widower is to deceased spouse as widow is to deceased spouse. Geography and Location:
Get the Full Details
Paris is to France as Tokyo is to Japan. London is to England as Rome is to Italy. Cairo is to Egypt as Lisbon is to Portugal. Sydney is to Australia as Toronto is to Canada. Berlin is to Germany as Madrid is to Spain. Moscow is to Russia as Beijing is to China. New York is to America as Ottawa is to Canada. Mumbai is to India as Seoul is to South Korea. Dubai is to UAE as Singapore is to Singapore Island. Reykjavik is to Iceland as Helsinki is to Finland. Body Parts: Eye is to see as ear is to hear. Hand is to grasp as foot is to walk. Lung is to breathe as heart is to pump. Liver is to filter as kidney is to purify. Skin is to protect as skull is to shield. Finger is to point as toe is to balance. Tongue is to taste as nose is to smell. Brain is to think as spinal cord is to transmit signals. Stomach is to digest as intestines are to absorb nutrients. Retina is to vision as cochlea is to hearing.
Animal Relationships: Cat is to kitten as dog is to puppy. Cow is to calf as horse is to foal. Chicken is to chick as duck is to duckling. Sheep is to lamb as goat is to kid. Bear is to cub as lion is to cub. Wolf is to pack as elephant is to herd. Fish is to school as birds are to flock. Bee is to hive as ants are to colony. Whale is to pod as deer is to herd. Eagle is to nest as spider is to web. Verbs and Actions:
Read is to book as watch is to movie. Sing is to voice as dance is to body. Write is to pen as paint is to brush. Run is to sprint as walk is to stroll. Eat is to devour as sip is to taste. Sleep is to dream as awake is to observe. Buy is to spend as sell is to earn. Give is to receive as take is to obtain. Ask is to question as answer is to respond. Begin is to start as finish is to complete. Materials and Objects: Wood is to tree as stone is to mountain. Cotton is to plant as wool is to sheep. Glass is to sand as steel is to iron ore. Rubber is to latex as paper is to pulp. Silk is to silkworm as leather is to cow. Bronze is to copper as brass is to zinc. Glass is to fragile as diamond is to hard. Iron is to metal as wood is to organic material. Silver is to jewelry as concrete is to construction. Water is to liquid as ice is to solid.
Units and Measurements: Meter is to length as gram is to weight. Second is to time as liter is to volume. Kelvin is to temperature as ampere is to current. Hertz is to frequency as newton is to force. Joule is to energy as watt is to power. Mole is to amount as candela is to luminous intensity. Kilometer is to distance as mile is to distance. Liter is to capacity as ounce is to weight. Kilogram is to mass as ton is to heavy mass. Volt is to voltage as ohm is to resistance. Food and Cooking:
Bread is to wheat as cheese is to milk. Coffee is to bean as tea is to leaf. Butter is to cream as sugar is to cane. Rice is to grain as pasta is to flour. Apple is to tree as carrot is to root. Steak is to beef as salmon is to fish. Chocolate is to cocoa as vanilla is to orchid. Salt is to mineral as pepper is to berry. Soup is to liquid as salad is to vegetable mix. Fruit is to dessert as appetizer is to starter. Nature and Weather: Rain is to wet as snow is to cold. Sun is to light as moon is to darkness. Wind is to breeze as hurricane is to storm. Cloud is to sky as wave is to ocean. River is to flow as mountain is to elevation. Forest is to trees as desert is to sand. Lightning is to thunder as sunrise is to dawn. Thunder is to storm as rainbow is to clear sky. Earthquake is to tremor as volcano is to eruption. Season is to year as month is to calendar week.
Technology and Tools: Computer is to software as engine is to fuel. Phone is to call as radio is to broadcast. Camera is to photo as printer is to document. Keyboard is to type as mouse is to navigate. Screen is to display as speaker is to audio. Battery is to charge as solar panel is to energy. Code is to program as blueprint is to building. Website is to domain as app is to platform. Server is to hosting as router is to connection. USB is to port as HDMI is to video output. Time and Calendar:
Second is to minute as minute is to hour. Hour is to day as day is to week. Week is to month as month is to year. Year is to decade as decade is to century. Century is to millennium as era is to geological period. Morning is to sunrise as evening is to sunset. Spring is to bloom as autumn is to harvest. Winter is to snow as summer is to heat. Past is to history as future is to prediction. Today is to present as yesterday is to memory. Social and Cultural: Friend is to companion as enemy is to opponent. Leader is to follow as follower is to obey. Guest is to host as tenant is to landlord. Stranger is to unknown as neighbor is to familiar. Citizen is to country as resident is to city. Hero is to brave as coward is to fearful. Teacher is to student as mentor is to protégé. Celebrity is to fame as recluse is to privacy. Immigrant is to relocate as native is to birthplace. Volunteer is to service as professional is to career.
Emotions and States: Happiness is to joy as sadness is to grief. Anger is to rage as fear is to terror. Calm is to peace as anxiety is to worry. Hope is to optimism as despair is to hopelessness. Love is to affection as hate is to loathing. Pride is to confidence as shame is to embarrassment. Curiosity is to wonder as boredom is to apathy. Confusion is to puzzlement as clarity is to understanding. Guilt is to remorse as relief is to comfort. Envy is to jealousy as contentment is to satisfaction. Clothing and Accessories:
Shirt is to upper body as pants are to lower body. Shoe is to foot as glove is to hand. Hat is to head as belt is to waist. Coat is to outerwear as scarf is to neck accessory. Sock is to foot as tie is to neck. Dress is to formal as t-shirt is to casual. Watch is to wrist as glasses are to face. Jacket is to arms as hoodie is to torso. Ring is to finger as bracelet is to wrist. Boot is to winter as sandal is to summer. Vehicles and Transport: Car is to road as boat is to water. Train is to track as airplane is to sky. Bicycle is to pedal as skateboard is to push. Bus is to passengers as taxi is to individual. Truck is to cargo as van is to transport. Helicopter is to hover as glider is to glide. Motorcycle is to rider as sidecar is to passenger. Ship is to ocean as submarine is to deep water. Ferry is to short crossing as cruise ship is to long voyage. Skateboard is to tricks as skateboard ramp is to stunt surface.
Plants and Gardening: Tree is to forest as flower is to garden. Root is to anchor as leaf is to photosynthesis. Seed is to plant as egg is to organism. Tree is to trunk as stem is to stalk. Flower is to bloom as fruit is to ripen. Weed is to unwanted as crop is to cultivated. Palm is to tropical as pine is to coniferous. Cactus is to desert as lotus is to swamp. Vine is to climb as bush is to spread. Herb is to seasoning as spice is to flavoring. Math and Science:
Plus is to add as minus is to subtract. Multiplication is to product as division is to quotient. Triangle is to three sides as square is to four sides. Atom is to molecule as cell is to organism. Gravity is to fall as magnetism is to attract. Density is to mass over volume as velocity is to displacement over time. Prime is to indivisible as composite is to divisible. Circle is to round as polygon is to many-sided. Hypothesis is to guess as theory is to proven framework. Kinetic is to motion as potential is to stored energy. Colors and Visual: Red is to stop as green is to go. Blue is to ocean as yellow is to sun. Black is to darkness as white is to light. Green is to nature as red is to danger. Purple is to royalty as gold is to wealth. Brown is to earth as gray is to concrete. Pink is to soft as orange is to vibrant. Cyan is to blue-green as magenta is to red-purple. Indigo is to rainbow as violet is to light spectrum end. White is to blank as black is to filled.
Numbers and Quantities: One is to single as two is to pair. Half is to 50 percent as quarter is to 25 percent. Double is to times two as triple is to times three. Even is to divisible by two as odd is to not divisible by two. Zero is to nothing as infinity is to endless. First is to begin as last is to conclude. Few is to small amount as many is to large quantity. All is to everything as none is to nothing. Some is to partial as every is to total. Less is to reduce as more is to increase. Daily Life and Routine:
Breakfast is to morning as dinner is to evening. Brush is to teeth as comb is to hair. Wake is to morning as sleep is to night. Work is to job as hobby is to leisure. Pay is to salary as bill is to expense. Lock is to secure as key is to unlock. Open is to begin as close is to end. Fill is to pour in as empty is to pour out. Clean is to remove dirt as dirty is to covered in grime. Fresh is to new as stale is to old and unused. Common Pitfalls with Analogies The main thing people miss is that word analogies depend entirely on training data quality. If your corpus lacks examples of a particular relationship, the vector math simply won't work. I ran into this when a client needed domain-specific analogies for a medical terminology project. The standard embeddings had no concept of "prescription is to pharmacist as diagnosis is to doctor." The model kept mapping it to unrelated terms because the training data never connected those concepts relationally.
The workaround was to fine-tune the embeddings on a specialized medical text corpus. That took about six hours on a decent GPU and produced dramatically better results. You can't fix a fundamentally broken embedding space with clever prompting. The relationships need to actually exist in the vector geometry. Another issue is that some analogies look correct but exploit spurious correlations. Models sometimes associate words based on co-occurrence rather than genuine semantic relationships. "Man is to king as woman is to queen" works because both pairs share similar context distributions. But "Man is to doctor as woman is to nurse" also produces a plausible result, even though that reflects societal bias in the training data more than any logical relationship. You have to be careful about what conclusions you draw from analogy scores. The real limitation is that word analogies only capture a narrow slice of meaning. They work well for categorical and relational knowledge but fail completely on context-dependent usage. A word like "run" has dozens of meanings depending on context, and analogies can't represent any of that nuance. If you need that level of understanding, you should be looking at transformer-based models with contextual embeddings rather than static word vectors. The analogy task was never meant to solve general language understanding. It was always a benchmark for embedding quality, and treating it as anything more is a mistake.
For most practical applications, I recommend using analogy tasks as a quick sanity check on your embedding model rather than a core feature. If your analogies score in the 70 to 80 percent range on standard benchmarks, your model is probably adequate for downstream tasks. If it scores below 50 percent, you either need more training data or a fundamentally different approach. There is no workaround for bad embeddings.
