You might want to consider the start of this tutorial.
Short introductions to other TF datasets:
- Dead Sea Scrolls,
- Old Babylonian Letters, or the
- Q'uran
%load_ext autoreload
%autoreload 2
from tf.app import use
VERSION = "2021"
A = use("etcbc/bhsa", hoist=globals())
Rough edges¶
It might be helpful to peek under the hood, especially when exploring searches that go slow.
If you went through the previous parts of the tutorial you have encountered cases where things come to a grinding halt.
Yet we can get a hunch of what is going on, even in those cases.
For that, we use the lower-level search api S of Text-Fabric, and not the
wrappers that the high level A api provides.
The main difference is, that S.search() returns a generator of the results,
whereas A.search() returns a list of the results.
In fact, A.search() calls the generator function delivered by S.search() as often as needed.
For some queries, the fetching of results is quite costly, so costly that we do not want to fetch
all results up-front. Rather we want to fetch a few, to see how it goes.
In these cases, directly using S.search() is preferred over A.search().
query = """
book
chapter
verse
phrase det=und
word lex=>LHJM/
"""
Study¶
First we call S.study(query).
The syntax will be checked, features loaded, the search space will be set up, narrowed down, and the fetching of results will be prepared, but not yet executed.
In order to make the query a bit more interesting, we lift the constraint that the results must be in Genesis 1-2.
S.study(query)
Before we rush to the results, lets have a look at the plan.
S.showPlan()
Here you see already what your results will look like.
Each result r is a tuple of nodes:
(R0, R1, R2, R3, R4)that instantiate the objects in your template.
In case you are curious, you can get details about the search space as well:
S.showPlan(details=True)
The part about the nodes shows you how many possible instantiations for each object in your template has been found. These are not results yet, because only combinations of instantiations that satisfy all constraints are results.
The constraints come from the relations between the objects that you specified.
In this case, there is only an implicit relation: embedding [[.
Later on we'll examine all
spatial relations.
The part about the edges shows you the constraints, and in what order they will be computed when stitching results together. In this case the order is exactly the order by which the relations appear in the template, but that will not always be the case. Text-Fabric spends some time and ingenuity to find out an optimal stitch plan. Fetching results is like selecting a node, stitching it to another node with an edge, and so on, until a full stitch of nodes intersects with all the node sets from which they must be chosen (the yarns).
Fetching results may take time.
For some queries, it can take a large amount of time to walk through all results. Even worse, it may happen that it takes a large amount of time before getting the first result. During stitching, many stitchings will be tried and fail before they can be completed.
This has to do with search strategies on the one hand, and the very likely possibility to encounter pathological search patterns, which have billions of results, mostly unintended. For example, a simple query that asks for 5 words in the Hebrew Bible without further constraints, will have 425,000 to the power of 5 results. That is 10-e28 (a one with 28 zeros), roughly the number of molecules in a few hundred liters of air. That may not sound much, but it is 10,000 times the amount of bytes that can be currently stored on the whole Internet.
Text-Fabric search is not yet done with finding optimal search strategies, and I hope to refine its arsenal of methods in the future, depending on what you report.
Counting results¶
It is always a good idea to get a feel for the amount of results, before you dive into them head-on.
S.count(progress=1, limit=5)
We asked for 5 results in total, with a progress message for every one. That was a bit conservative.
S.count(progress=100, limit=500)
Still pretty quick, now we want to count all results.
S.count(progress=200, limit=None)
Fetching results¶
It is time to see something of those results.
S.fetch(limit=10)
Not very informative.
Just a quick observation: look at the last column.
These are the result nodes for the word part in the query, indicated as R7 by showPlan() before.
And indeed, they are all below 425,000, the number of words in the Hebrew Bible.
Nevertheless, we want to glean a bit more information off them.
for r in S.fetch(limit=10):
print(S.glean(r))
Caution¶
It is not possible to do
len(S.fetch()). Becausefetch()is a generator, not a list. It will deliver a result every time it is being asked and for as long as there are results, but it does not know in advance how many there will be.Fetching a result can be costly, because due to the constraints, a lot of possibilities may have to be tried and rejected before a the next result is found.
That is why you often see results coming in at varying speeds when counting them.
We can also use A.table() to make a list of results.
This function is part of the Bhsa API, not of the generic Text-Fabric machinery, as opposed to S.glean().
So, you can use S.glean() for every Text-Fabric corpus, but the output is still not very nice.
A.table() gives much nicer output.
A.table(S.fetch(limit=5))
Queries with abundant results¶
Above we mentioned that there are queries with astronomically many results. Here we present one:
query = """
word
# word
"""
We are asking for any pair of different words. That will give roughly 425,000 * 425,000 results, which is 180 billion results. This is a lot to produce, it will take time on even the best of computers, and once you've got the results, what would you do with them. Let's see what happens if we count these results.
S.study(query)
S.count(progress=500000)
Text-fabric has cut off the process at a certain limit. This limit is a number of times the maximum node in your corpus:
5787324 / F.otype.maxNode
If you really need more results than this limit, you can specify a higher limit:
S.count(progress=500000, limit=8 * F.otype.maxNode)
Now you do not get an error message, because you got what you've asked for.
Or, in the advanced interface, let's fetch the standard maximum of results:
results = A.search(query)
Or, with a modified limit:
results = A.search(query, limit=5 * F.otype.maxNode)
Again, you do not get an error message, because you got what you've asked for.
Slow queries¶
The search template above has some pretty tight constraints on one of its objects, so the amount of data to deal with is pretty limited.
If the constraints are weak, search may become slow.
For example, here is a query that looks for pairs of phrases in the same clause in such a way that one is engulfed by the other.
query = """
% test
% verse book=Genesis chapter=2 verse=25
verse
clause
p1:phrase
w1:word
w3:word
w1 < w3
p2:phrase
w2:word
w1 < w2
w3 > w2
p1 < p2
"""
A couple of remarks you may have encountered before.
- some objects have got a name
- there are additional relations specified between named objects
<means: comes before, and>: comes after in the canonical order for nodes, which for words means: comes textually before/after, but for other nodes the meaning is explained here- later on we describe those relations in more detail
Note on order Look at the words
w1andw3below phrasep1. Although in the templatew1comes beforew3, this is not translated in a search constraint of the same nature.Order between objects in a template is never significant, only embedding is.
Because order is not significant, you have to specify order yourself, using relations.
It turns out that this is better than the other way around.
In MQL order is significant, and it is very difficult to
search for w1 and w2 in any order.
Especially if your are looking for more than 2 complex objects with lots of feature
conditions, your search template would explode if you had to spell out all
possible permutations. See the example of Reinoud Oosting below.
Note on gaps Look at the phrases
p1andp2. We do not specify an order here, only that they are different. In order to prevent duplicated searches withp1andp2interchanged, we even stipulate thatp1 < p2. There are many spatial relationships possible between different objects. In many cases, neither the one comes before the other, nor vice versa. They can overlap, one can occur in a gap of the other, they can be completely disjoint and interleaved, etc.
# ignore this
# S.tweakPerformance(yarnRatio=2)
S.study(query)
Text-Fabric knows that narrowing down the search space in this case would take ages, without resulting in a significantly shrunken space. So it skips doing so for most constraints.
Let us see the plan, with details.
S.showPlan(details=True)
As you see, we have a hefty search space here.
Let us play with the count() function.
S.count(progress=10, limit=100)
We can be bolder than this!
S.count(progress=100, limit=1000)
OK, not too bad, but note that it takes a big fraction of a second to get just 100 results.
Now let us go for all of them by the thousand.
S.count(progress=1000, limit=None)
See? This is substantial work.
A.table(S.fetch(limit=5))
Hand-coding¶
As a check, here is some code that looks for basically the same phenomenon: a phrase within the gap of another phrase. It does not use search, and it gets a bit more focused results, in half the time compared to the search with the template.
Hint If you are comfortable with programming, and what you look for is fairly generic, you may be better off without search, provided you can translate your insight in the data into an effective procedure within Text-Fabric. But wait till we are completely done with this example!
TF.indent(reset=True)
TF.info("Getting gapped phrases")
results = []
for v in F.otype.s("verse"):
for c in L.d(v, otype="clause"):
ps = L.d(c, otype="phrase")
first = {}
last = {}
slots = {}
# make index of phrase boundaries
for p in ps:
words = L.d(p, otype="word")
first[p] = words[0]
last[p] = words[-1]
slots[p] = set(words)
for p1 in ps:
for p2 in ps:
if p2 < p1:
continue
if len(slots[p1] & slots[p2]) != 0:
continue
if first[p1] < first[p2] and last[p2] < last[p1]:
results.append(
(v, c, p1, p2, first[p1], first[p2], last[p2], last[p1])
)
TF.info("{} results".format(len(results)))
Pretty printing¶
We can use the pretty printing of A.table() and A.show() here as well, even though we have
not used search!
Not that you can show the node numbers. In this case it helps to see where the gaps are.
A.table(results, withNodes=True, end=5)
A.show(results, start=1, end=1)
NB Gaps are a tricky phenomenon. In gaps we will deal with them cruelly.
Performance tuning¶
Here is an example by Yanniek van der Schans (2018-09-21).
query = """
c:clause
PreGap:phrase_atom
LastPhrase:phrase_atom
:=
Gap:clause_atom
:: word
PreGap < Gap
Gap < LastPhrase
c || Gap
"""
Here are the current settings of the performance parameters:
S.tweakPerformance()
S.study(query)
S.showPlan(details=True)
S.count(progress=1, limit=3)
Can we do better?
The performance parameter yarnRatio can be used to increase the amount of preprocessing, and we can
increase to number of random samples that we make by tryLimitFrom and tryLimitTo.
We start with increasing the amount of up-front edge-spinning.
S.tweakPerformance(yarnRatio=0.2, tryLimitFrom=10000, tryLimitTo=10000)
S.study(query)
S.showPlan(details=True)
It seems to be the same plan.
S.count(progress=1, limit=3)
No improvement.
What if we decrease the amount of edge spinning?
S.tweakPerformance(yarnRatio=5, tryLimitFrom=10000, tryLimitTo=10000)
S.study(query)
S.showPlan(details=True)
S.count(progress=1, limit=3)
Again, no improvement.
We'll look for queries where the parameters matter more in the future.
Here is how to reset the performance parameters:
S.tweakPerformance(yarnRatio=None, tryLimitFrom=None, tryLimitTo=None)
All steps¶
- start your first step in mastering the bible computationally
- display become an expert in creating pretty displays of your text structures
- search turbo charge your hand-coding with search templates
advanced sets relations quantifiers fromMQL rough
You have seen cases where the implementation is to blame.
Now I want to point to gaps in your understanding:
- exportExcel make tailor-made spreadsheets out of your results
- share draw in other people's data and let them use yours
- export export your dataset as an Emdros database
- annotate annotate plain text by means of other tools and import the annotations as TF features
- map map somebody else's annotations to a new version of the corpus
- volumes work with selected books only
- trees work with the BHSA data as syntax trees
CC-BY Dirk Roorda