What a dialect map is trying to show
A dialect is a variety of a language associated with a region or social group, with its own typical pronunciation, vocabulary and sometimes grammar. Traditional dialect research relied on fieldworkers who visited places, interviewed local speakers and recorded their answers to set questions, such as what people called a particular object. Those surveys produced the familiar maps with lines marking where one word gives way to another.
Fieldwork gives careful, high-quality data, but it is slow and costly, and it covers few speakers. Social media offers a very different source: huge volumes of informal writing, much of it by ordinary people, with some information about where it came from. That has made online posts attractive to researchers interested in how regional dialects vary and change.
How online writing is used
The typical approach goes roughly as follows.
- Collect posts. Researchers gather a large set of public messages, often restricted to those that carry location information.
- Pick features. They choose words, spellings or constructions that might vary by region, such as different terms for the same thing or distinctive spellings that mimic local pronunciation.
- Count by place. For each area, they measure how often the feature appears relative to all the text from that area.
- Map and cluster. The results are plotted, and statistical methods may group areas with similar patterns together.
Findings from such work have often matched expectations from traditional surveys, which is reassuring. In some cases they have also pointed to regional vocabulary that older surveys had not covered, including newer words or forms that spread through informal writing. They can show how a term moves between cities, and whether language patterns follow geography, population centres or links between places.
Where location data comes from, and why it is shaky
The location attached to a post is rarely a clean measure of where the writer grew up. Several kinds of location signal exist, and each has issues.
- Precise device location. Only some users switch it on, and those who do may differ from those who do not.
- A profile location field. People may leave it blank, enter a joke, or list a place they identify with rather than where they live.
- Inferred location. Systems may guess from text, friends or network connections, which introduces circular reasoning when the aim is to study text and place together.
Even a correct location may mislead. Someone posting from a city may have grown up elsewhere, so their dialect reflects their childhood home. Dialect is generally tied to where a person learned the language, not where they happen to be. Students, workers who have moved and travellers can all blur the picture, especially in large cities with high mobility.
Who is writing
The people who post publicly are not a random sample of the population, which raises the familiar issue of sampling frames. Users of a given platform often skew younger and more urban than the general population, and differ in education, income and interests. Older speakers, who may hold the most traditional forms of a dialect, are often underrepresented. Different platforms attract different crowds, so results from one may not apply to another.
Posting rates also vary by region. If a handful of very active accounts produce much of the text from a small town, one person’s habits can look like a local trend. Careful studies try to limit the weight of any single account, but this adjustment is not always done or described.
Writing is not speech
Dialect is mostly about speaking, and writing is a filtered version. People who write informally often use spelling to mimic speech, for example by altering letters to show an accent. Those choices are partly deliberate performance, and different writers may represent the same sound in different ways, or not at all. A person who speaks with a strong local accent may write in standard spelling, which hides their dialect from a text-based method. Others may write in a mock dialect for humour.
Vocabulary and grammar are more reliably recorded than pronunciation, but even then, people adjust their style to the audience. Writing to friends differs from writing in public, and public posts are the ones researchers can see.
Other ways the method can mislead
Trending topics and events. A local event, a sports team or a news story can cause a burst of regional words that reflects the topic rather than dialect. Names of places and teams, in particular, can dominate a regional comparison unless removed.
Bots and repeated content. Automated or duplicate posts can inflate a feature’s apparent frequency in a place. The wider concern over fake accounts applies here as well.
Sparse data. For small areas, there may be too few posts to say anything with confidence. Maps that show smooth colour gradients can hide that some regions are based on very little text.
Platform changes. When a service changes how location is recorded or who can see posts, the data for a region can change without any change in language.
Over-reading differences. A difference between two places in how often a word appears is not necessarily a boundary between dialects. Words vary along gradients, and individual features rarely line up neatly into a single dialect area. Choosing different features can produce different maps.
What a cautious reader should look for
Studies and news stories about online dialect maps are most trustworthy when they explain the following: how location was determined, how many posts and how many distinct writers sit behind each area, how the researchers handled events and place names, and whether the findings were checked against another source such as interviews or recordings.
Used this way, social media data is a valuable supplement to traditional dialect work, strong at showing broad patterns in informal vocabulary and fast changes. It is a poor substitute for listening to people speak, and it says more about the language of particular online users than about everyone who lives in a region. A map built from online posts should therefore be read as a picture of a certain kind of writing, by a certain kind of poster, in places linked to them by an imperfect signal.