swh:1:snp:7ce5f1105410d5ee1ad6abfdc873986c25b579e5
Tip revision: 207dd83ba491914b8bb76346c79d98cf3ac74366 authored by Dirk Roorda on 24 January 2024, 21:27:55 UTC
small fix in NER layout css
small fix in NER layout css
Tip revision: 207dd83
prepare.py
"""
# Pre-compute data.
For TF to work efficiently, some derived data needs to be pre-computed.
The pre-computed data has a similar function as indexes in a database.
Pre-computation is triggered when `tf.fabric.Fabric` loads features, and
the order and nature of the steps is configured in
`tf.core.fabric.PRECOMPUTE`.
The functions in this module implement those tasks.
"""
import collections
import functools
import array
from .helpers import itemize
def levels(info, error, otype, oslots, otext):
"""Computes level data.
For each node type, compute the average number of slots occupied by its nodes,
and order the node types on that.
Parameters
----------
info: function
Method to write informational messages to the console.
error: function
Method to write error messages to the console.
otype: iterable
The data of the `otype` feature.
oslots: iterable
The data of the `oslots` feature.
otext: iterable
The data of the `otext` feature.
Returns
-------
tuple
An ordered tuple, each member with the information of a node type:
* node type name
* average number of slots contained in the nodes of this type
* first node of this type
* last node of this type
The order of the tuple is descending by average number of slots per node of that
type.
Notes
-----
!!! explanation "Level computation and customization"
All node types have a level, defined by the average amount of slots object of
that type usually occupy. The bigger the average object, the lower the levels.
Books have the lowest level, words the highest level.
However, this can be overruled. Suppose you have a node type `phrase` and above
it a node type `cluster`, i.e. phrases are contained in clusters, but not vice
versa. If all phrases are contained in clusters, and some clusters have more
than one phrase, the automatic level ranking of node types works out well in
this case. But if clusters only have very small phrases, and the big phrases do
not occur in clusters, then the algorithm may assign a lower rank to clusters
than to phrases.
In general, it is too expensive to try to compute the levels in a sophisticated
way. In order to remedy cases where the algorithm assigns wrong levels, you can
add a `@levels` and / or `@levelsConstraint` key to the `otext`
configuration feature.
See `tf.core.text`.
"""
(otype, maxSlot, maxNode, slotType) = otype
oslots = oslots[0]
levelOrder = otext.get("levels", None)
if levelOrder is not None:
levelRank = {level: i for (i, level) in enumerate(levelOrder.split(","))}
otypeCount = collections.Counter()
otypeMin = {}
otypeMax = {}
slotSetLengths = collections.Counter()
info("get ranking of otypes")
for k in range(len(oslots)):
ntp = otype[k]
otypeCount[ntp] += 1
slotSetLengths[ntp] += len(oslots[k])
tfn = k + maxSlot + 1
if ntp not in otypeMin:
otypeMin[ntp] = tfn
if ntp not in otypeMax or otypeMax[ntp] < tfn:
otypeMax[ntp] = tfn
sortKey = (
(lambda x: -x[1])
if levelOrder is None
else (lambda x: (-1, levelRank[x[0]]) if x[0] in levelRank else (1, -x[1]))
)
result = sorted(
(
(
ntp,
slotSetLengths[ntp] / otypeCount[ntp],
otypeMin[ntp],
otypeMax[ntp],
)
for ntp in otypeCount
),
key=sortKey,
) + [(slotType, 1, 1, maxSlot)]
resultIndex = {r[0]: i for (i, r) in enumerate(result)}
levelConstraintsSpec = otext.get("levelConstraints", None)
if levelConstraintsSpec:
levelConstraints = levelConstraintsSpec.split(";")
for levelConstraint in levelConstraints:
(smaller, biggerSpec) = levelConstraint.strip().split("<")
smaller = smaller.strip()
biggers = {b.strip() for b in biggerSpec.strip().split(",")}
highestBigIndex = max(resultIndex.get(tp, 0) for tp in biggers)
smallerIndex = resultIndex.get(smaller, len(result))
if smallerIndex <= highestBigIndex:
x = result.pop(smallerIndex)
result.insert(highestBigIndex, x)
resultIndex = {r[0]: i for (i, r) in enumerate(result)}
info("results:")
for (otp, av, omin, omax) in result:
info(f"{otp:<15}: {round(av, 2):>8} {{{omin}-{omax}}}", tm=False)
return tuple(result)
def order(info, error, otype, oslots, levels):
"""Computes order data for the canonical ordering.
The canonical ordering between nodes is defined in terms of the slots that
nodes contain, and if that is not decisive, the rank of the node type is taken
into account, and if that is still not decisive, the node itself is taken into
account.
Parameters
----------
info: function
Method to write informational messages to the console.
error: function
Method to write error messages to the console.
otype: iterable
The data of the `otype` feature.
oslots: iterable
The data of the `oslots` feature.
levels: tuple
The data of the `levels` pre-computation step.
Returns
-------
tuple
All nodes, slot and nonslot, in canonical order.
See Also
--------
tf.core.nodes: canonical ordering
"""
(otype, maxSlot, maxNode, slotType) = otype
oslots = oslots[0]
info("assigning otype levels to nodes")
otypeLevels = dict(((x[0], i) for (i, x) in enumerate(reversed(levels))))
def otypeRank(n):
return otypeLevels[slotType if n < maxSlot + 1 else otype[n - maxSlot - 1]]
def before(na, nb):
if na < maxSlot + 1:
a = na
sa = {a}
else:
a = na - maxSlot
sa = set(oslots[a - 1])
if nb < maxSlot + 1:
b = nb
sb = {b}
else:
b = nb - maxSlot
sb = set(oslots[b - 1])
oa = otypeRank(na)
ob = otypeRank(nb)
if sa == sb:
return (
(-1 if na < nb else 1 if na > nb else 0)
if oa == ob
else -1
if oa > ob
else 1
)
if sa > sb:
return -1
if sa < sb:
return 1
am = min(sa - sb)
bm = min(sb - sa)
return -1 if am < bm else 1 if bm < am else 0
canonKey = functools.cmp_to_key(before)
info("sorting nodes")
nodes = sorted(range(1, maxNode + 1), key=canonKey)
# return array.array("I", nodes)
return tuple(nodes)
def rank(info, error, otype, order):
"""Computes rank data.
The rank of a node is its place in among the other nodes in the
canonical order (see `tf.core.nodes`).
Parameters
----------
info: function
Method to write informational messages to the console.
error: function
Method to write error messages to the console.
otype: iterable
The data of the `otype` feature.
order: tuple
The data of the `order` feature.
Returns
-------
tuple
The ranks of all nodes, slot and nonslot, with respect to the canonical order.
"""
(otype, maxSlot, maxNode, slotType) = otype
info("ranking nodes")
nodesRank = dict(((n, i) for (i, n) in enumerate(order)))
return array.array("I", (nodesRank[n] for n in range(1, maxNode + 1)))
# return tuple((nodesRank[n] for n in range(1, maxNode + 1)))
def levUp(info, error, otype, oslots, rank):
"""Computes level-up data.
Level-up data is used by the API function `tf.core.locality.Locality.u`.
This function computes the embedders of a node by looking them up from
the level-up data.
Parameters
----------
info: function
Method to write informational messages to the console.
error: function
Method to write error messages to the console.
otype: iterable
The data of the `otype` feature.
oslots: iterable
The data of the `oslots` feature.
rank: tuple
The data of the `rank` pre-computation step.
Returns
-------
tuple
The `n`-th member is a tuple of the embedder nodes of `n`.
Those tuples are sorted in canonical order (`tf.core.nodes`).
Notes
-----
!!! hint "Memory efficiency"
Many nodes have the same tuple of embedders.
Those embedder tuples will be reused for those nodes.
Warnings
--------
It is not advisable to use this data directly by `C.levUp.data`,
it is far better to use the `tf.core.locality.Locality.u` function.
Only when every bit of performance waste has to be squeezed out,
this raw data might be a deal.
"""
(otype, maxSlot, maxNode, slotType) = otype
oslots = oslots[0]
info("making inverse of edge feature oslots")
oslotsInv = {}
for (k, mList) in enumerate(oslots):
for m in mList:
oslotsInv.setdefault(m, set()).add(k + 1 + maxSlot)
info("listing embedders of all nodes")
embedders = []
for n in range(1, maxSlot + 1):
contentEmbedders = oslotsInv.get(n, tuple())
embedders.append(
tuple(
sorted(
(m for m in contentEmbedders if m != n),
key=lambda k: -rank[k - 1],
)
)
)
for n in range(maxSlot + 1, maxNode + 1):
mList = oslots[n - maxSlot - 1]
if len(mList) == 0:
embedders.append(tuple())
else:
contentEmbedders = functools.reduce(
lambda x, y: x & oslotsInv[y],
mList[1:],
oslotsInv[mList[0]],
)
embedders.append(
tuple(
sorted(
(m for m in contentEmbedders if m != n),
key=lambda k: -rank[k - 1],
)
)
)
# reuse embedder tuples, because lots of nodes share embedders
seen = {}
embeddersx = []
for t in embedders:
if t not in seen:
# seen[t] = array.array("I", t)
seen[t] = tuple(t)
embeddersx.append(seen[t])
return tuple(embeddersx)
def levDown(info, error, otype, levUp, rank):
"""Computes level-down data.
Level-down data is used by the API function `tf.core.locality.Locality.d`.
This function computes the embedded nodes of a node by looking them up from
the level-down data.
Parameters
----------
info: function
Method to write informational messages to the console.
error: function
Method to write error messages to the console.
otype: iterable
The data of the `otype` feature.
levUp: iterable
The data of the `levUp` pre-computation step.
rank: tuple
The data of the `rank` pre-computation step.
Returns
-------
tuple
The `n`-th member is an tuple of the embedded nodes of `n + maxSlot`.
Those tuples are sorted in canonical order (`tf.core.nodes`).
!!! hint "Memory efficiency"
Slot nodes do not have embedded nodes, so they do not have to occupy
space in this tuple. Hence the first member are the embedded nodes
of node `maxSlot + 1`.
!!! caution "Use with care"
It is not advisable to use this data directly by `C.levDown.data`,
it is far better to use the `tf.core.locality.Locality.d` function.
Only when every bit of performance waste has to be squeezed out,
this raw data might be a deal.
"""
(otype, maxSlot, maxNode, slotType) = otype
info("inverting embedders")
inverse = {}
for n in range(maxSlot + 1, maxNode + 1):
for m in levUp[n - 1]:
inverse.setdefault(m, set()).add(n)
info("turning embeddees into list")
embeddees = []
for n in range(maxSlot + 1, maxNode + 1):
embeddees.append(
array.array("I", sorted(inverse.get(n, []), key=lambda m: rank[m - 1]))
# tuple(sorted(inverse.get(n, []), key=lambda m: rank[m - 1]))
)
return tuple(embeddees)
def characters(info, error, otext, tFormats, *tFeats):
"""Computes character data.
For each text format, a frequency list of the characters in that format
is made.
Parameters
----------
info: function
Method to write informational messages to the console.
error: function
Method to write error messages to the console.
otext: iterable
The data of the `otext` feature.
tFormats: dict
Dictionary keyed by text format and valued by the tuple of features
used in that format.
tFeats: iterable
Each `tFeat` is the name and the data of a text feature.
i.e. a feature used in text formats.
Returns
-------
dict
Keyed by format valued by a frequency dict, which is
itself keyed by single characters and valued by the frequency
of that character in the whole corpus when rendered with that format.
"""
charFreqsByFeature = {}
for (tFeat, data) in tFeats:
freqList = collections.Counter()
if data is not None:
for v in data.values():
freqList[v] += 1
charFreq = collections.defaultdict(lambda: 0)
for (v, freq) in freqList.items():
for c in str(v):
charFreq[c] += freq
charFreqsByFeature[tFeat] = charFreq
charFreqsByFmt = {}
for (fmt, tFeatures) in sorted(tFormats.items()):
charFreq = collections.defaultdict(lambda: 0)
for tFeat in tFeatures:
thisCharFreq = charFreqsByFeature[tFeat]
for (c, freq) in thisCharFreq.items():
charFreq[c] += freq
charFreqsByFmt[fmt] = sorted(x for x in charFreq.items())
return charFreqsByFmt
def boundary(info, error, otype, oslots, rank):
"""Computes boundary data.
For each slot, the nodes that start at that slot and the nodes that end
at that slot are collected.
Boundary data is used by the API functions
`tf.core.locality.Locality.p`.
and
`tf.core.locality.Locality.n`.
Parameters
----------
info: function
Method to write informational messages to the console.
error: function
Method to write error messages to the console.
otype: iterable
The data of the `otype` feature.
oslots: iterable
The data of the `oslots` feature.
rank: tuple
The data of the `rank` pre-computation step.
Returns
-------
tuple
* first: tuple of tuple
The `n`-th member is the tuple of nodes that start at slot `n`,
ordered in *reversed* canonical order (`tf.core.nodes`);
* last: tuple of tuple
The `n`-th member is the tuple of nodes that end at slot `n`,
ordered in canonical order;
Notes
-----
!!! hint "why reversed canonical order?"
Just for symmetry.
"""
(otype, maxSlot, maxNode, slotType) = otype
oslots = oslots[0]
firstSlotsD = {}
lastSlotsD = {}
for (node, slots) in enumerate(oslots):
realNode = node + 1 + maxSlot
firstSlotsD.setdefault(slots[0], []).append(realNode)
lastSlotsD.setdefault(slots[-1], []).append(realNode)
firstSlots = tuple(
tuple(sorted(firstSlotsD.get(n, []), key=lambda node: -rank[node - 1]))
# array.array("I", sorted(firstSlotsD.get(n, []),
# key=lambda node: -rank[node - 1]))
for n in range(1, maxSlot + 1)
)
lastSlots = tuple(
tuple(sorted(lastSlotsD.get(n, []), key=lambda node: rank[node - 1]))
# array.array("I", sorted(lastSlotsD.get(n, []),
# key=lambda node: rank[node - 1]))
for n in range(1, maxSlot + 1)
)
return (firstSlots, lastSlots)
def sections(info, error, otype, oslots, otext, levUp, levels, *sFeats):
"""Computes section data.
TF datasets may define up to three section levels, roughly corresponding
with a volume, a chapter, a paragraph.
If the corpus has a richer section structure, it is also possible
a different, more flexible and more extensive nest of structural sections.
See `structure`.
TF must be able to go from sections at one level to the sections
at one level lower. It must also be able to map section headings
to nodes. For this, the section features are needed, since they
contain the section headings.
Parameters
----------
info: function
Method to write informational messages to the console.
error: function
Method to write error messages to the console.
otype: iterable
The data of the `otype` feature.
oslots: iterable
The data of the `oslots` feature.
otext: iterable
The data of the `otext` feature.
levUp: tuple
The data of the `levUp` pre-computation step.
levels: tuple
The data of the `levels` pre-computation step.
sFeats: iterable
Each `sFeat` is the data of a section feature.
Returns
-------
tuple
* `sec1`
Mapping from section-level-1 nodes to mappings from
section-level-2 headings to section-level-2 nodes.
* `sec2`
Mapping from section-level-1 nodes to mappings from
section-level-2 headings to mappings from
section-level-3 headings to section-level-3 nodes.
Warnings
--------
Note that the terms `book`, `chapter`, `verse` are not baked into TF.
It is the corpus data, especially the `otext` configuration feature that
spells out the names of the sections.
"""
(otype, maxSlot, maxNode, slotType) = otype
oslots = oslots[0]
support = {level[0]: (level[2], level[3]) for level in levels}
sTypes = itemize(otext["sectionTypes"], ",")
sec1 = {}
sec2 = {}
nestingProblems = collections.Counter()
if len(sTypes) < 2:
pass
elif len(sTypes) < 3:
c1 = 0
support1 = support[sTypes[1]]
for n1 in range(support1[0], support1[1] + 1):
n0s = tuple(x for x in levUp[n1 - 1] if otype[x - maxSlot - 1] == sTypes[0])
if not n0s:
nestingProblems[
f"section {sTypes[1]} without containing {sTypes[0]}"
] += 1
continue
n0 = n0s[0]
n1head = sFeats[1].get(n1, None)
if n1head is None:
nestingProblems[f"{sTypes[1]}-node {n1} has no section heading"] += 1
if n0 not in sec1:
sec1[n0] = {}
if n1head not in sec1[n0]:
sec1[n0][n1head] = n1
c1 += 1
info(f"{c1} {sTypes[1]}s indexed")
else:
c1 = 0
c2 = 0
support2 = support[sTypes[2]]
for n2 in range(support2[0], support2[1] + 1):
n0s = tuple(x for x in levUp[n2 - 1] if otype[x - maxSlot - 1] == sTypes[0])
if not n0s:
nestingProblems[
f"section {sTypes[2]} without containing {sTypes[0]}"
] += 1
continue
n0 = n0s[0]
n1s = tuple(x for x in levUp[n2 - 1] if otype[x - maxSlot - 1] == sTypes[1])
if not n1s:
nestingProblems[
f"section {sTypes[2]} without containing {sTypes[1]}"
] += 1
continue
n1 = n1s[0]
n1head = sFeats[1].get(n1, None)
if n1head is None:
nestingProblems[f"{sTypes[1]}-node {n1} has no section heading"] += 1
n2head = sFeats[2].get(n2, None)
if n2head is None:
nestingProblems[f"{sTypes[2]}-node {n2} has no section heading"] += 1
if n0 not in sec1:
sec1[n0] = {}
if n1head not in sec1[n0]:
sec1[n0][n1head] = n1
c1 += 1
sec2.setdefault(n0, {}).setdefault(n1head, {})[n2head] = n2
c2 += 1
info(f"{c1} {sTypes[1]}s and {c2} {sTypes[2]}s indexed")
if nestingProblems:
for (msg, amount) in sorted(nestingProblems.items()):
error(f"WARNING: {amount:>4} x {msg}")
return (sec1, sec2)
def structure(info, error, otype, oslots, otext, rank, levUp, *sFeats):
"""Computes structure data.
If the corpus has a rich section structure, it is possible to define
a flexible and extensive nest of structural sections.
Independent of this,
TF datasets may also define up to three section levels,
roughly corresponding with a volume, a chapter, a paragraph.
See `sections`.
TF must be able to go from sections at one level to the sections
at one level lower. It must also be able to map section headings
to nodes. For this, the section features are needed, since they
contain the section headings.
Parameters
----------
info: function
Method to write informational messages to the console.
error: function
Method to write error messages to the console.
otype: iterable
The data of the `otype` feature.
oslots: iterable
The data of the `oslots` feature.
otext: iterable
The data of the `otext` feature.
rank: tuple
The data of the `rank` pre-computation step.
levUp: tuple
The data of the `levUp` pre-computation step.
sFeats: iterable
Each `sFeat` the data of a section feature.
Returns
-------
tuple
* `headingFromNode` (Mapping from nodes to section keys)
* `nodeFromHeading` (Mapping from section keys to nodes)
* `multiple`
* `top`
* `up`
* `down`
Notes
-----
A section key of a structural node is obtained by going a level up from
that node, retrieving the heading of that structural node, then going up again,
and so on till a top node is reached. The tuple of headings obtained in this way
is the section key.
"""
(otype, maxSlot, maxNode, slotType) = otype
oslots = oslots[0]
sTypeList = itemize(otext["structureTypes"], ",")
nsTypes = len(sTypeList)
nsFeats = len(sFeats)
if nsTypes != nsFeats:
error(
f"WARNING: {nsTypes} structure levels but {nsFeats} corresponding features"
)
return ({}, {})
sTypes = set(sTypeList)
if len(sTypes) != nsTypes:
error("WARNING: duplicate structure levels")
return ({}, {})
higherTypes = collections.defaultdict(set)
for (i, highType) in enumerate(sTypeList):
for lowType in sTypeList[i:]:
higherTypes[lowType].add(highType)
featFromType = {sTypeList[i]: sFeats[i] for i in range(nsTypes)}
multiple = collections.defaultdict(list)
headingFromNode = {}
nodeFromHeading = {}
for n in range(maxSlot + 1, maxNode + 1):
nType = otype[n - maxSlot - 1]
if nType not in sTypes:
continue
ups = (u for u in levUp[n - 1] if otype[u - maxSlot - 1] in higherTypes[nType])
sKey = tuple(
reversed(
tuple(
(
otype[x - maxSlot - 1],
featFromType[otype[x - maxSlot - 1]].get(x, None),
)
for x in (n, *ups)
)
)
)
if sKey in nodeFromHeading:
if sKey not in multiple:
multiple[sKey].append(nodeFromHeading[sKey])
multiple[sKey].append(n)
nodeFromHeading[sKey] = n
headingFromNode[n] = sKey
multiple = {
sKey: tuple(sorted(ns, key=lambda n: rank[n - 1]))
for (sKey, ns) in multiple.items()
}
top = tuple(
sorted(
(n for (n, h) in headingFromNode.items() if len(h) == 1),
key=lambda n: rank[n - 1],
)
)
up = {}
for (n, heading) in headingFromNode.items():
lHeading = len(heading)
if lHeading == 1:
continue
upNode = None
for i in range(lHeading - 1, 0, -1):
upHeading = heading[0:i]
upNode = nodeFromHeading.get(upHeading, None)
if upNode is not None:
up[n] = upNode
break
down = {}
for (n, heading) in headingFromNode.items():
if len(heading) == 1:
continue
down.setdefault(up[n], []).append(n)
down = {n: tuple(sorted(ms, key=lambda m: rank[m - 1])) for (n, ms) in down.items()}
return (headingFromNode, nodeFromHeading, multiple, top, up, down)
