City of Fredericton data projectCan a computer tell from a description, in English or French, whether a Fredericton amenity is open all year?
The City of Fredericton lists its recreation amenities, from playgrounds and ball diamonds to dog parks, with a description in English, a description in French and a season. Our students hide the season and train a naive Bayes classifier on the words. All accuracy figures here are our own results; the descriptions and seasons are the City’s.
Our naive Bayes results on City of Fredericton Recreation Amenities (283 amenities tagged Summer or Year-Round, fetched 22 October 2026; each one tested while the other 282 are used for training)| Model | What it reads | Share of amenities whose season it guesses right |
|---|
| Baseline | Nothing: always says Year-Round, the commoner season | 54.1% |
| Naive Bayes, English | The words of the English description | 88.7% |
| Naive Bayes, French | The words of the French description | 85.2% |
This is the complete program, using only Python’s standard library. It downloads the amenity list, turns each description into a set of lower-case words, and for every amenity in turn trains on all the others and guesses the left-out one, a method called leave-one-out that suits a small dataset. The add-one in the formula keeps a word the model has never seen with a season from zeroing out the whole calculation. Its output on 22 October 2026 is the three-row table above:import json, math, re, urllib.request
from collections import Counter
URL = ("https://services2.arcgis.com/iLWAxhpxafhOza2U/arcgis/rest/services/"
"Recreation_Amenities_view/FeatureServer/0/query?where=1%3D1"
"&outFields=Description,DescriptionFR,Season&returnGeometry=false&f=json")
rows = [f["attributes"] for f in json.load(urllib.request.urlopen(URL))["features"]]
rows = [r for r in rows if r["Season"] in ("Summer", "Year-Round")]
def words(text):
return set(re.findall(r"\w+", (text or "").lower()))
def train_and_guess(train, sample):
seasons = Counter(season for _, season in train)
counts = {s: Counter() for s in seasons}
for bag, season in train:
counts[season].update(bag)
vocab = set().union(*(bag for bag, _ in train))
def score(s):
total = sum(counts[s].values())
return math.log(seasons[s] / len(train)) + sum(
math.log((counts[s][w] + 1) / (total + len(vocab))) for w in sample if w in vocab)
return max(seasons, key=score)
def leave_one_out(field):
data = [(words(r[field]), r["Season"]) for r in rows]
right = sum(train_and_guess(data[:i] + data[i + 1:], bag) == season
for i, (bag, season) in enumerate(data))
return round(100 * right / len(data), 1)
top = Counter(r["Season"] for r in rows).most_common(1)[0]
print(len(rows), "amenities; always", repr(top[0]), "is right",
round(100 * top[1] / len(rows), 1), "% of the time")
print("naive Bayes on English descriptions:", leave_one_out("Description"), "%")
print("naive Bayes on French descriptions:", leave_one_out("DescriptionFR"), "%")
1. Words as evidence
Naive Bayes treats each word as a small piece of evidence. In the English descriptions, “diamond” and “ball” appear 25 times each, always on summer amenities, while “playground” appears 57 times on year-round amenities and only twice on summer ones.
2. Surprising clues
The word “drop”, from “drop-in”, appears 31 times, every time on a summer amenity. The model has no idea what drop-in means; it only learned that the word travels with summer. That is the strength and the weakness of statistical learning.
3. English or French?
The French descriptions give 85.2 per cent against 88.7 in English. Students investigate why: French joins words with apostrophes, uses more forms of each word, and spreads the same evidence over more spellings. Handling that is real natural language processing.
Where the model fails, and why that is useful
The English model makes 32 mistakes. Seven are summer parking areas called year-round and four are year-round parking areas called summer: a parking description says little about the season, and the City lists 28 parking rows as year-round and 11 as summer. Six more are community school gyms. When a model fails on one category again and again, the honest conclusion is that the text does not contain the answer, not that a cleverer model is needed.
A few notes on the data. Of the 295 amenities in the layer, 11 are tagged Winter and one has no season, so the model compares only Summer (130) and Year-Round (153). The descriptions were written by people, not generated, which is why identical amenities often share a template sentence and why the model does well on them. The dataset sits in the City’s Fredericton Open Data catalogue, published under the Open Government License - City of Fredericton, version 1.0, which the City says is based on the Open Government Licence - Canada and whose grant extends to use “including for commercial purposes”. Source: City of Fredericton, Recreation Amenities, City of Fredericton open data. Nothing here carries the City’s endorsement.
