Keith Briggs

 home · papers · thesis · talks · meetings · records · maths notes · software · languages · music · travel · cv · students · maps · place-names « · people · photos · links · ex libris · site map

### Bayesian analysis of UK place-names - revised 2008 version

or, where is Ambridge, anyway?

#### The idea

This is an idea inspired by these initially disconnected thoughts:

• Drawing maps of the distribution of place-name types is an old game - see e.g. Smith's English place-name elements or Watts' Cambridge dictionary of English place-names (pages li-lxii). I have even automated the process myself with my maps here. But this process is deterministic.
• By contrast, we all have a feeling about different place-name types in the various regions. For example, Buckfastleigh could only be in Devon, Giggleswick in Yorkshire, and Flegg in Norfolk. But is there any science in this? Could we capture feelings such as these in a mathematical model? If we could, it would have to be a probabilistic model.
• In the last few years big advances in Bayesian text classification (both in theory and in practice) have been stimulated by the requirements of spam filtering. Can we exploit these advances?
• Made-up place-names in novels and films always sound wrong, even though they are made up of the right elements. Why is this?

#### The method

So this is what I have done:

• I built Bayesian models for each square in the National Grid, using all village, town, farm, river etc. names in each square from a gazetteer of 251307 names in total (my previous version used only 29326 names).
• For each non-existent name below, I computed the posterior probability for it to be in each square according to the model. Some of my made-up names now actually exist (Chopton, Hanton, Marsley), so I have made up slightly different new names in these cases (Chipton, Hanniton, Maresley).
• If the probability was above 1%, I coloured the square with an approximately spectral colour proportional to the probability. Dark blue indicates a very small probability, and bright red the maximum probability, with the levels in between being light blue, green, yellow, and orange, except that in the maps, the largest probability is scaled to 1 to determine the colour. See the key lower down on this page.
• Thus, the reddest square would be the likeliest place for the non-existent name to be located, if it really existed.
• In essence, I am doing text classification with 49 categories, rather than the three (spam, possible spam, legitimate email) as is done in spam filtering.

#### Some initial checks

Just to check that the whole idea is sensible, we'll do some real names first. Note that the model does not know the location of any place-name. But if this idea is at all reasonable, it should get real English place-names in roughly the right places, and it should be able to locate Welsh and Gaelic names. Here are some checks of this type. (Note that there really is a Cambridge in Gloucestershire.)

Here, the lower the peak probability, the less the name resembles a real name.

#### Technical details

This is the colour coding used in the maps:

For computing the Bayesian models, I used the dbacl software by Laird Breyer.

The next table gives the number of names per square used to build the Bayesian models, and the easting, northing, latitude and longitude of its bottom left-hand corner.

squarenumbereastingnorthinglatitudelongitude
NA 143 0 900 57.810 -8.739
NB 2569 100 900 57.889 -7.063
NC 3426 200 900 57.945 -5.380
ND 1723 300 900 57.978 -3.691
NF 1724 0 800 56.918 -8.577
NG 4732 100 800 56.994 -6.941
NH 6243 200 800 57.048 -5.298
NJ 8340 300 800 57.080 -3.650
NM 5303 100 700 56.098 -6.825
NN 6639 200 700 56.151 -5.220
NO 7906 300 700 56.182 -3.611
NR 3795 100 600 55.203 -6.716
NS 9727 200 600 55.253 -5.147
NT 8673 300 600 55.284 -3.575
NU 1129 400 600 55.294 -2.000
NW 91 100 500 54.307 -6.613
NX 5954 200 500 54.356 -5.078
NY 11883 300 500 54.386 -3.540
NZ 5803 400 500 54.395 -2.000
SD 9264 300 400 53.487 -3.507
SE 11233 400 400 53.496 -2.000
SH 6325 200 300 52.561 -4.951
SJ 12063 300 300 52.588 -3.476
SK 9911 400 300 52.597 -2.000
SN 8551 200 200 51.663 -4.892
SO 11753 300 200 51.689 -3.447
SP 9163 400 200 51.698 -2.000
SS 6278 200 100 50.764 -4.836
ST 10496 300 100 50.790 -3.419
SU 10435 400 100 50.799 -2.000
SW 2960 100 0 49.824 -6.172
SX 6787 200 0 49.866 -4.783
SY 2052 300 0 49.891 -3.392
SZ 1146 400 0 49.900 -2.000
TA 1847 500 400 53.487 -0.493
TF 5051 500 300 52.588 -0.524
TG 1642 600 300 52.561 0.951
TL 9503 500 200 51.689 -0.553
TM 4453 600 200 51.663 0.892
TQ 10194 500 100 50.790 -0.581
TR 1550 600 100 50.764 0.836