Reconstructing Pompeian Households

David Mimno
Department of Computer Science
University of Massachusetts, Amherst
Amherst, MA 01003
[email protected]

Abstract

A database of objects discovered in houses in the Roman city of Pompeii provides

a unique view of ordinary life in an ancient city. Experts have used this collec-
tion to study the structure of Roman households, exploring the distribution and
variability of tasks in architectural spaces, but such approaches are necessarily af-
fected by modern cultural assumptions. In this study we present a data-driven ap-
proach to household archeology, treating it as an unsupervised labeling problem,
that attempts to provide a more objective complement to human interpretation.

1 Introduction

Over the past century the goal of archeology has shifted from finding objects of artistic value to re-
constructing the details of day-to-day life in ancient cultures. Reconstructing daily life is difficult as
the evidence that would be useful was generally considered commonplace and unworthy of preser-
vation by people at the time. Objects are gradually moved or disposed of; structures are renovated,
repurposed or recycled for building materials. These ongoing processes make it difficult to make
statements about specific points in time based on material evidence. A remarkable exception to this
pattern is the Roman city of Pompeii in southern Italy, which was was covered in volcanic ash during
the eruption of Mt. Vesuvius in 79 AD. As a result of its sudden, violent destruction, many aspects
of day-to-day life in first-century Roman Italy were preserved in place as they existed at a specific
moment in time.
The study of Pompeian households has until recently been dominated by analyses of architectural
patterns and wall paintings, as these are easily available for study. Only in the past few years
has a database of objects, which are removed for conservation immediately upon excavation, been
compiled and made available by Allison [1]. This database, which is available online,1 contains more
than 6000 artifact records for finds in 30 architecturally similar “atrium-style” houses in Pompeii.
Artifacts are annotated with one of 240 typological categories (coin, amphora, etc.) and which of
the 574 rooms they were found in. Allison has used data about artifacts in their original context
to challenge many common assumptions about the function of particular types of object, the use of
particular spaces, and the consistency of patterns of use across different houses.
The relatively large amount of archeological information compiled in the Pompeian households
database supports the application of statistical data mining tools. In this paper we apply one such
tool, statistical topic modeling [2], in which rooms are modeled as having mixtures of functions,
and functions are modeled as distributions over a “vocabulary” of object types. The purpose of this
study is not to show that topic modeling is the best tool for archeological investigation, but that it
is an appropriate tool that can provide a complement to human analysis. To this aim, we attempt
to provide a perspective on several issues raised by Allison, that is, if not unbiased, then at least
mathematically concrete in its biases.
1
http://www.stoa.org/projects/ph/home

1
2 Context-driven estimation of artifact function

When artifacts are excavated, standard archeological practice involves removing them to secure
storage for preservation. Although the location of artifacts is carefully noted in modern scientific
digs, artifacts in storage tend to be analyzed in comparison to typologically similar objects rather
than within their original context. As a result, questions about the use or function of particular
artifact types have been based more on fairly arbitrary tradition and researchers’ perceptions of
what an artifact is reminiscent of than the types of objects found with the object. For example,
Allison identifies two classes of artifact, the casseruola (“casserole dish”) and forma di pasticceria
(“pastry mold”), that were named based on similarities to 19th century household objects. She
cites research going back more than 100 years suggesting that these items were not used in food
preparation contexts, but finds that contemporary scholars still make this assumption due to their
(modern) names.
In order to reduce bias of this kind, we explore the function of these artifact types using only cooc-
currence data, without any reference to the actual typology of the objects. We reduce all 240 object
descriptions to integers, and then apply a statistical topic model to detect clusters of object cooc-
currence that may indicate functions. Note that we are still completely dependent on experts to
classify physical objects into appropriate categories, but given those classifications we make no
further archeological assumptions in training the model.
Table 2 shows object distributions from a single model with 20 “topics” or “functional groups.”
Functional groups are presented in descending order by average number of artifacts in left to right
order. We train the model using Gibbs sampling, with 50,000 iterations, saving samples every 200
iterations.
Intuitively, if two objects share a similar pattern of use, they should both have reasonably high
probability in one or more “topic”. Given a model, we can evaluate the probability that two ob-
ject types xP and y will be produced by the same functional group, marginalized over functions, as
P (x, y) = t P (t)P (x | t)P (y | t), where P (t) is proportional to the average number of tokens
assigned to functional group t and P (x | t) is proportional to the average number of objects of type
x assigned to functional group t. Table 1 shows results for the two types mentioned previously.
There is little to no connection to food preparation objects, supporting Allison’s claim that modern
names for these items are incorrect.

Table 1: Joint probability of two controversial object types with other object types. Objects most likely to
occur with bronze casseruola, marginalized over 20 functional groups (“topics”) are shown on the left: the first
item solidly associated with food preparation is well down the list. Similarly, objects likely to occur with a
forma di pasticceria are shown on the right, again showing no significant connection to food preparation.

bronze casseruola silver vessel/forma di pasticceria

0.00036 door/chest/cupboard fitting 0.00008 jewelry
0.00035 glass bottle/flask/pyxis 0.00004 silver cup/bowl/cup fragment
0.00028 small glass bottle 0.00004 chest/cupboard fitting
0.00026 pottery jug 0.00004 silver patera/casseruola/plate
0.00024 bronze jug/jug fragment 0.00003 casket fitting
0.00022 chest/cupboard fitting 0.00002 coin
0.00020 ceramic lamp 0.00002 bronze jug/jug fragment
0.00018 jewelry 0.00002 door/chest/cupboard fitting
0.00016 pottery beaker/small vase 0.00002 bronze or silver spoon
0.00015 coin 0.00002 chest fitting
0.00013 pottery pot 0.00001 box/casket
... 0.00001 part of coin hoard
0.00010 table/table fittings/table base 0.00001 small glass bottle
0.00010 pottery jar/vase 0.00001 hair pin
0.00010 bronze cooking pot/basin/pot/fragment 0.00001 ceramic lamp
(a) (b)

2
Table 2: 20 “topics” derived from objects (tokens) in rooms (documents) in Pompeian households.

0.212 coin 0.193 door/chest/cupboard fitting

0.086 ceramic lamp 0.088 chest/cupboard fitting
0.070 small glass bottle 0.067 ring
0.061 door/chest/cupboard fitting 0.062 coin
0.058 jewelry 0.057 weights
0.033 architectural fitting 0.054 nail
0.165 pottery amphora/amphoretta/hydria 0.566 chest/cupboard fitting
0.120 pottery amphora fragment/lid 0.046 glass bottle/flask/pyxis
0.103 unidentified pottery vessel 0.043 chest/cista
0.095 nail 0.038 chest fitting
0.086 pottery jar/vase 0.028 cupboard
0.061 ceramic lamp 0.021 scales
0.194 glass bottle/flask/pyxis 0.142 pottery jug
0.116 small glass bottle 0.090 pottery pot
0.060 glass beaker/cup 0.069 ceramic lamp
0.055 small glass jar/vase 0.048 terra sigillata bowl/cup
0.054 jewelry 0.039 pottery cup/small bowl
0.042 small bronze/wooden container 0.035 bronze casseruola
0.184 pottery amphora/amphoretta/hydria 0.161 recess
0.126 building material 0.120 niche
0.106 stairway 0.118 hearth/focolare
0.076 impluvium/compluvium 0.114 latrine
0.060 puteal/puteal fragment 0.093 stairway
0.051 table/table fittings/table base 0.071 drain/pipe/tap/cess pit
0.176 bronze jug/jug fragment 0.323 part of coin hoard
0.070 bronze/iron lamp/base/fragment 0.212 jewelry
0.068 lampstand/lampstand fragment/candelabrum 0.205 human skeleton
0.047 bronze casseruola 0.028 architectural door fitting
0.043 amphora 0.019 bed/couch
0.037 bronze serpent-handled basin 0.017 bag
0.403 architectural door fitting 0.359 large sculpture/sculpture fragment
0.376 door/chest/cupboard fitting 0.073 fixed statue base
0.031 chest/cupboard fitting 0.068 fountain/fountain fitting
0.023 fixed seat 0.059 marble or stone basin
0.020 key 0.056 marble base/statue base/basin base
0.018 nail 0.048 table/table fittings/table base
0.368 ceramic lamp 0.199 jewelry
0.053 small glass bottle 0.113 silver cup/bowl/cup fragment
0.052 weights 0.102 silver patera/casseruola/plate/plate fragment
0.049 other statuary/sculpture/fragments 0.071 casket fitting
0.048 pottery beaker/small vase 0.045 chest/cupboard fitting
0.036 glass plate/tray 0.041 bronze or silver spoon
0.424 bed/couch fragment 0.137 pottery jar/vase
0.068 large sculpture/sculpture fragment 0.136 bronze cooking pot/basin/pot/fragment
0.062 bed/couch 0.066 hearth/focolare
0.061 marble slab/tondo/decorated fragment 0.066 tripod
0.031 door/chest/cupboard fitting 0.065 pottery basin/krater
0.024 pottery amphora/amphoretta/hydria 0.064 pottery large bowl
0.288 cupboard fitting 0.745 shelving/mezzanine/suspension nails
0.085 chest/cupboard fitting 0.024 recess
0.085 miscellaneous metal object 0.021 pottery amphora/amphoretta/hydria
0.075 cupboard 0.013 pottery pot
0.064 unidentified furniture 0.007 built-in cupboard
0.047 glass jar/vase 0.007 bronze/iron lamp/base/fragment
0.076 measuring equipment 0.355 vehicle fragment
0.050 iron or lead strip 0.309 harness
0.043 chisel 0.065 pendant
0.034 pick/pickaxe 0.029 coin
0.033 bone fitting/strip 0.022 cart/wagon
0.033 hook or knife 0.015 small glass bottle

3
3 Modeling room functions
The Roman houses included in the database show strong architectural patterns. The street entrance
usually opens onto an large room with an opening in the roof. This “atrium” connects to an open
garden area. Both large rooms are surrounded by smaller rooms. Much of the study of Pompeian
households has involved identifying categories of rooms and assigning functions to them. Again,
Allison argues that commonly held assumptions about such functions are incorrect. For example, the
atrium is generally described as a formal space, in which the pater familias received his clients and
distributed gifts from a large metal chest. Allison argues that the atrium was more of a utilitarian,
industrial space. Similarly, a closed room off the atrium is usually described as a cubiculum (“bed-
room”), but there is little evidence that these rooms correspond to modern concepts of bedrooms.
Before using an analysis of objects to provide information about the function of spaces, it is im-
portant to establish whether architectural room types have consistent patterns of object contents.
The city suffered a severe earthquake 17 years before the final eruption, and was disrupted before
and after the eruption. We use a likelihood ratio test to compare the hypothesis that the contents
of all rooms of a particular architectural type are drawn from a single multinomial distribution over
functional groups against the hypothesis that each room has its own multinomial. The log of this
likelihood ratio is positive, indicating support for the “single multinomial” hypothesis, for rooms
with clear functions such as entryways, kitchens, and bath complexes. The log ratio is most neg-
ative, indicating greater variability, for the atrium and garden areas. These results suggest that we
should be careful in drawing conclusions. In addition, we applied a Dirichlet process clustering
algorithm but did not find consistent clusters.
We can, however, attempt to rule out certain possibilities. Table 3 shows probable functional groups
for several architectural types. Atria show evidence of utilitarian storage: large ceramic amphorae.
The only type that shows significant evidence of objects related to bedding is type 11 (large open
rooms with views of the garden); the cubiculum does not.

Table 3: Distributions of functional groups in common room types

Type 3 (atrium)
0.279 pottery amphora/amphoretta/hydria, building material, stairway, impluvium/compluvium, puteal/puteal fragment
0.108 chest/cupboard fitting, chest/cista, glass bottle/flask/pyxis, chest fitting, cupboard
0.083 coin, ceramic lamp, small glass bottle, architectural fitting, jewelry
0.077 pottery jug, ceramic lamp, table/table fittings/table base, terra sigillata bowl/cup, seashell/conch/snail shell

Type 4 (cubiculum)
0.122 recess, built-in cupboard, stairway, niche, unidentified fixture/mound
0.116 shelving/mezzanine/suspension nails, recess, pottery amphora/amphoretta/hydria, cistern head, pottery pot
0.092 coin, ceramic lamp, small glass bottle, architectural fitting, jewelry
0.065 pottery jug, pottery amphora/amphoretta/hydria, pottery pot, pottery jar/vase, pottery plate/dish/tray

Type 11 (exedra, oecus, triclinium)

0.111 door/chest/cupboard fitting, coin, ring, chest/cupboard fitting, nail
0.109 recess, built-in cupboard, stairway, niche, unidentified fixture/mound
0.080 bed/couch fragment, door/chest/cupboard fitting, large sculpture/sculpture fragment, architectural door fitting, bed/couch
0.058 human skeleton, jewelry, part of coin hoard, bed/couch, bag

References
[1] P. Allison. Pompeian households: an analysis of the material culture. Cotsen Institute of
Archaeology, 2001.
[2] D. Blei, A. Ng, and M. Jordan. Latent Dirichlet allocation. JMLR, 2003.