diff --git a/treecat/tables.py b/treecat/tables.py
index ae59d68..3ec123f 100644
--- a/treecat/tables.py
+++ b/treecat/tables.py
@@ -6,13 +6,22 @@
 
 import numpy as np
 
+from treecat.util import TODO
+
 TY_MULTINOMIAL = 0
+TY_CATEGORICAL = 1
+TY_ORDINAL = 2
 
 
 class Table(object):
     """Internal representation of read-only row-oriented heterogeneous data."""
 
-    def __init__(self, feature_types, ragged_index, data):
+    def __init__(
+            self,
+            feature_types,
+            ragged_index,
+            data,
+            categorical_data=None, ):
         """Create a read-only table of heterogeneous data.
 
         Args:
@@ -21,24 +30,47 @@ def __init__(self, feature_types, ragged_index, data):
                 data array.
             data: An [N, _]-shaped numpy array of ragged data, where the vth
                 column is stored in data[:, ragged_index[v]:ragged_index[v+1]].
+            categorical_data: An [N, Vc]-shaped numpy array of category
+                indices, where Vc is the number of categorical features and N
+                is the number fo rows.
         """
+        N = data.shape[0]
+        V = len(feature_types)
+
         feature_types = np.asarray(feature_types, dtype=np.int8)
         feature_types.flags.writeable = False
         assert len(feature_types.shape) == 1
 
+        # Multinomial features.
         ragged_index = np.asarray(ragged_index, np.int32)
         ragged_index.flags.writeable = False
         assert len(ragged_index.shape) == 1
-        assert ragged_index.shape[0] == 1 + feature_types.shape[0]
-
+        assert ragged_index.shape[0] == 1 + V
         data = np.asarray(data, np.int8)
         data.flags.writeable = False
         assert len(data.shape) == 2
+        assert data.shape[0] == N
         assert data.shape[1] == ragged_index[-1]
 
+        # Categorical features.
+        if categorical_data is None:
+            categorical_data = np.empty([N, 0], dtype=np.int16)
+        categorical_data = np.asarray(categorical_data, dtype=np.int16)
+        assert len(categorical_data.shape) == 2
+        assert categorical_data.shape[0] == N
+
         self._feature_types = feature_types
-        self._ragged_index = ragged_index
-        self._data = data
+        self._multinomial_ragged_index = ragged_index
+        self._multinomial_data = data
+        self._categorical_data = categorical_data
+
+    @property
+    def num_rows(self):
+        return self._multinomial_data.shape[0]
+
+    @property
+    def num_cols(self):
+        return self._feature_types.shape[0]
 
     @property
     def feature_types(self):
@@ -46,16 +78,34 @@ def feature_types(self):
 
     @property
     def ragged_index(self):
-        return self._ragged_index
+        return self._multinomial_ragged_index
 
     @property
     def data(self):
-        return self._data
+        return self._multinomial_data
 
     @property
-    def num_rows(self):
-        return self._data.shape[0]
+    def categorical_data(self):
+        return self._categorical_data
 
-    @property
-    def num_cols(self):
-        return self._feature_types.shape[0]
+
+_INTERNALIZE = {
+    TY_MULTINOMIAL: TY_MULTINOMIAL,
+    TY_CATEGORICAL: TY_MULTINOMIAL,
+    TY_ORDINAL: TY_MULTINOMIAL,
+}
+
+
+def internalize_table(table):
+    """Transform features of a table to types with internal support.
+
+    Args:
+        table: A Table object with arbitrary feature types.
+
+    Returns:
+        A Table object with transformed feature types.
+    """
+    feature_types = np.array([_INTERNALIZE[ty]
+                              for ty in table.feature_types], np.int8)
+    for v, (old, new) in enumerate(zip(table.feature_types, feature_types)):
+        TODO()