Christian,

Adding diacritics sensitive slows execution by a factor of 3. My script (fragment below), which joins two large databases, namely CT.gov http://clinicaltrials.gov and DrugBank, takes 2 hours without the diacritics sensitive constraint but 6 hours with it. Given the combinatorics involved, I am wondering if there is a better way to do this in BaseX.

Thanks, Ron

for $drug in db:open('DrugBank')/drugbank/drug let $drug_name := $drug/name/text() let $drug_synonyms := functx:value-union(normalize-space(lower-case($drug/name)), local:drug-synonyms($drug_name)) for $synonym_name in $drug_synonyms ... for $study in db:open('CTGov')/clinical_study[intervention/intervention_name contains text { $synonym_name } using case insensitive using diacritics sensitive] ...

Ron Katriel, Ph.D. | Principal Data Scientist | Medidata Solutions http://www.mdsol.com/ 350 Hudson Street, 7th Floor, New York, NY 10014 rkatriel@mdsol.com tbrophy@mdsol.com | direct: +1 201 337 3622 <//201 337 3622> | mobile: +1 201 675 5598 <//+1 201 675 5598> | main: +1 212 918 1800 <//+1 212 918 1800>

On August 1, 2018 at 12:41:26 PM, Ron Katriel (rkatriel@mdsol.com) wrote:

Thanks, Christian. Strange, prior to contacting you and on a hunch, I tried adding the missing “using” keyword but still got the syntax error. Anyway, everything is good now!

Best, Ron

On August 1, 2018 at 3:57:51 AM, Christian Grün (christian.gruen@gmail.com) wrote:

I have fixed the example in the doc. Best, Christian

On Wed, Aug 1, 2018 at 5:08 AM Ron Katriel rkatriel@mdsol.com wrote:

Hi,

The following from your website (docs.basex.org/wiki/Full-Text) appears

to be syntactically incorrect

"'Äpfel' will not be found..." contains text "Apfel" diacritics sensitive

In the BaseX GUI the keyword diacritics is underlined in red and the

following error is reported

Unexpected end of query: 'diacritic sens...'.

This happens in version 8.6.4 and also the latest (9.0.2).

Thanks, Ron

Ron Katriel, Ph.D. | Principal Data Scientist | Medidata Solutions

350 Hudson Street, 7th Floor, New York, NY 10014

rkatriel@mdsol.com | direct: +1 201 337 3622 | mobile: +1 201 675 5598 |

main: +1 212 918 1800

Hi Gerrit,

Thanks for the suggestions. I would like to retain the original diacritics (for output purposes) but only match them when warranted (e.g., match acétazolamide to acétazolamide, but not acétazolamide to acetazolamide). I am looking for a simple solution that does not involve modifying the database or maintaining multiple copies (both for processing simplicity and storage efficiency reasons).

Thanks, Ron

On August 3, 2018 at 9:08:19 AM, Imsieke, Gerrit, le-tex ( gerrit.imsieke@le-tex.de) wrote:

Hi Ron,

You can add an extra element (or attribute) to the content when importing or modifying it. (Or another document in another database if you like – you can create and later find such an index document by giving it the same db:path as the original document.)

In this extra database, document, element and/or attribute, you can recreate the original text, except that you normalize the characters with diacritical marks to a canonical decomposition form and then strip away the diacritical marks like this:

replace(normalize-unicode($input, 'NFKD'), '\p{Mn}', '')

The full updating statement is beyond my cursory XQuery capabilities – I’d probably do it in XSLT. Also I don’t know how to trigger an event that would cause an update of the auxiliary fields when the underlying data changes.

Gerrit

On 03.08.2018 14:39, Ron Katriel wrote:

Christian,

Adding diacritics sensitive slows execution by a factor of 3. My script (fragment below), which joins two large databases, namely CT.gov <

https://urldefense.proofpoint.com/v2/url?u=http-3A__clinicaltrials.gov&d... and DrugBank, takes 2 hours without the

diacritics sensitive constraint but 6 hours with it. Given the combinatorics involved, I am wondering if there is a better way to do this in BaseX.

Thanks, Ron

for $drug in db:open('DrugBank')/drugbank/drug let $drug_name := $drug/name/text() let $drug_synonyms := functx:value-union(normalize-space(lower-case($drug/name)), local:drug-synonyms($drug_name)) for $synonym_name in $drug_synonyms ... for $study in db:open('CTGov')/clinical_study[intervention/intervention_name contains text { $synonym_name } using case insensitive using diacritics sensitive] ...

Ron Katriel, Ph.D. | Principal Data Scientist | Medidata Solutions http://www.mdsol.com/ 350 Hudson Street, 7th Floor, New York, NY 10014 rkatriel@mdsol.com mailto:tbrophy@mdsol.com | direct: +1 201 337 3622 tel://201%20337%203622 | mobile: +1 201 675 5598 tel://+1%20201%20675%205598 | main: +1 212 918 1800 tel://+1%20212%20918%201800

On August 1, 2018 at 12:41:26 PM, Ron Katriel (rkatriel@mdsol.com mailto:rkatriel@mdsol.com) wrote:

...

Thanks, Christian. Strange, prior to contacting you and on a hunch, I tried adding the missing “using” keyword but still got the syntax error. Anyway, everything is good now!

Best, Ron

On August 1, 2018 at 3:57:51 AM, Christian Grün (christian.gruen@gmail.com mailto:christian.gruen@gmail.com) wrote:

...

I have fixed the example in the doc. Best, Christian

On Wed, Aug 1, 2018 at 5:08 AM Ron Katriel <rkatriel@mdsol.com mailto:rkatriel@mdsol.com> wrote:

...

Hi,

The following from your website (docs.basex.org/wiki/Full-Text

<

https://urldefense.proofpoint.com/v2/url?u=http-3A__docs.basex.org_wiki_Full...) appears to be syntactically

...

...

incorrect

...

"'Äpfel' will not be found..." contains text "Apfel" diacritics

sensitive

...

...

...

In the BaseX GUI the keyword diacritics is underlined in red and the

following error is reported

...

...

...

Unexpected end of query: 'diacritic sens...'.

This happens in version 8.6.4 and also the latest (9.0.2).

Thanks, Ron

Ron Katriel, Ph.D. | Principal Data Scientist | Medidata Solutions

350 Hudson Street, 7th Floor, New York, NY 10014

rkatriel@mdsol.com mailto:rkatriel@mdsol.com | direct: +1 201 337

3622 | mobile: +1 201 675 5598 | main: +1 212 918 1800

...

Hi Christian,

Yes, I created a full-text index when the databases where loaded (see the commands below). I also verified that FTINDEX is true for both databases (in the GUI under Database > Open & Manage).

How do I ensure that my query is rewritten for index access?

Thanks, Ron

SET FTINDEX true; SET TOKENINDEX true; CREATE DB CTGov "/Data Sets/ ct.gov/xml" SET FTINDEX true; SET TOKENINDEX true; SET STRIPNS true; CREATE DB DrugBank “/Data Sets/DrugBank/drugbank.xml"

On August 3, 2018 at 4:12:43 PM, Christian Grün (christian.gruen@gmail.com) wrote:

Hi Ron,

Did you a) create a full-text index for your data and b) ensure that your query is rewritten for index access?

Best, Christian

On Fri, Aug 3, 2018 at 2:39 PM Ron Katriel rkatriel@mdsol.com wrote:

Christian,

Adding diacritics sensitive slows execution by a factor of 3. My script

(fragment below), which joins two large databases, namely CT.gov and DrugBank, takes 2 hours without the diacritics sensitive constraint but 6 hours with it. Given the combinatorics involved, I am wondering if there is a better way to do this in BaseX.

Thanks, Ron

for $drug in db:open('DrugBank')/drugbank/drug let $drug_name := $drug/name/text() let $drug_synonyms :=

functx:value-union(normalize-space(lower-case($drug/name)), local:drug-synonyms($drug_name))

for $synonym_name in $drug_synonyms ... for $study in

db:open('CTGov')/clinical_study[intervention/intervention_name contains text { $synonym_name } using case insensitive using diacritics sensitive]

...

Ron Katriel, Ph.D. | Principal Data Scientist | Medidata Solutions 350 Hudson Street, 7th Floor, New York, NY 10014 rkatriel@mdsol.com | direct: +1 201 337 3622 | mobile: +1 201 675 5598 |

main: +1 212 918 1800

On August 1, 2018 at 12:41:26 PM, Ron Katriel (rkatriel@mdsol.com) wrote:

Thanks, Christian. Strange, prior to contacting you and on a hunch, I

tried adding the missing “using” keyword but still got the syntax error. Anyway, everything is good now!

Best, Ron

On August 1, 2018 at 3:57:51 AM, Christian Grün (christian.gruen@gmail.com)

wrote:

I have fixed the example in the doc. Best, Christian

On Wed, Aug 1, 2018 at 5:08 AM Ron Katriel rkatriel@mdsol.com wrote:

...

Hi,

The following from your website (docs.basex.org/wiki/Full-Text) appears

to be syntactically incorrect

...

"'Äpfel' will not be found..." contains text "Apfel" diacritics

sensitive

...

In the BaseX GUI the keyword diacritics is underlined in red and the

following error is reported

...

Unexpected end of query: 'diacritic sens...'.

This happens in version 8.6.4 and also the latest (9.0.2).

Thanks, Ron

Ron Katriel, Ph.D. | Principal Data Scientist | Medidata Solutions

350 Hudson Street, 7th Floor, New York, NY 10014

rkatriel@mdsol.com | direct: +1 201 337 3622 | mobile: +1 201 675 5598

| main: +1 212 918 1800

...

Christian,

Thanks for sharing that. I assumed all along that this happens automatically. Anyway, I ran my query (for one drug, to save time) and see the following in the Info view

- apply text index for "Lenalidomide"

I believe the slow execution may be due to a combinatorial issue: the cross product of 280,000 clinical trials and ~10,000 drugs in DrugBank (not counting synonyms).

I am considering an algorithmic solution that involves storing the DrugBank information in a hash table (map) and looking it up while iterating through the CT.gov http://clinicaltrials.gov trials.

Best, Ron

On August 3, 2018 at 5:49:30 PM, Christian Grün (christian.gruen@gmail.com) wrote:

Our documentation should help you here: http://docs.basex.org/wiki/Indexes https://urldefense.proofpoint.com/v2/url?u=http-3A__docs.basex.org_wiki_Indexes&d=DwMFaQ&c=fi2D4-9xMzmjyjREwHYlAw&r=44jDQvzmnB_-ovfO6Iusj0ItciJrcWMOQQwd2peEBBE&m=mk1COTV1sAZu82fBqU9P70ZPQXi-d6NrV1-5QYTPHOo&s=Esza6Q3FyaDERIFJTWBAjifLIDVFW3bWKMLS4hbqv_A&e=

Ron Katriel rkatriel@mdsol.com schrieb am Fr., 3. Aug. 2018, 23:20:

Hi Christian,

Yes, I created a full-text index when the databases where loaded (see the commands below). I also verified that FTINDEX is true for both databases (in the GUI under Database > Open & Manage).

How do I ensure that my query is rewritten for index access?

Thanks, Ron

SET FTINDEX true; SET TOKENINDEX true; CREATE DB CTGov "/Data Sets/ ct.gov/xml https://urldefense.proofpoint.com/v2/url?u=http-3A__ct.gov_xml&d=DwMFaQ&c=fi2D4-9xMzmjyjREwHYlAw&r=44jDQvzmnB_-ovfO6Iusj0ItciJrcWMOQQwd2peEBBE&m=mk1COTV1sAZu82fBqU9P70ZPQXi-d6NrV1-5QYTPHOo&s=nDUqSutsQr7QyD8E6-XysRp1qudWO6I05tJaWjkCUI4&e= " SET FTINDEX true; SET TOKENINDEX true; SET STRIPNS true; CREATE DB DrugBank “/Data Sets/DrugBank/drugbank.xml"

On August 3, 2018 at 4:12:43 PM, Christian Grün (christian.gruen@gmail.com) wrote:

Hi Ron,

Did you a) create a full-text index for your data and b) ensure that your query is rewritten for index access?

Best, Christian

On Fri, Aug 3, 2018 at 2:39 PM Ron Katriel rkatriel@mdsol.com wrote:

...

Christian,

Adding diacritics sensitive slows execution by a factor of 3. My script

(fragment below), which joins two large databases, namely CT.gov and DrugBank, takes 2 hours without the diacritics sensitive constraint but 6 hours with it. Given the combinatorics involved, I am wondering if there is a better way to do this in BaseX.

...

Thanks, Ron

for $drug in db:open('DrugBank')/drugbank/drug let $drug_name := $drug/name/text() let $drug_synonyms :=

functx:value-union(normalize-space(lower-case($drug/name)), local:drug-synonyms($drug_name))

...

for $synonym_name in $drug_synonyms ... for $study in

db:open('CTGov')/clinical_study[intervention/intervention_name contains text { $synonym_name } using case insensitive using diacritics sensitive]

...

...

Ron Katriel, Ph.D. | Principal Data Scientist | Medidata Solutions 350 Hudson Street, 7th Floor, New York, NY 10014 rkatriel@mdsol.com | direct: +1 201 337 3622 | mobile: +1 201 675 5598

| main: +1 212 918 1800

...

On August 1, 2018 at 12:41:26 PM, Ron Katriel (rkatriel@mdsol.com)

wrote:

...

Thanks, Christian. Strange, prior to contacting you and on a hunch, I

tried adding the missing “using” keyword but still got the syntax error. Anyway, everything is good now!

...

Best, Ron

On August 1, 2018 at 3:57:51 AM, Christian Grün (

christian.gruen@gmail.com) wrote:

...

I have fixed the example in the doc. Best, Christian

On Wed, Aug 1, 2018 at 5:08 AM Ron Katriel rkatriel@mdsol.com wrote:

...

Hi,

The following from your website (docs.basex.org/wiki/Full-Text

https://urldefense.proofpoint.com/v2/url?u=http-3A__docs.basex.org_wiki_Full-2DText&d=DwMFaQ&c=fi2D4-9xMzmjyjREwHYlAw&r=44jDQvzmnB_-ovfO6Iusj0ItciJrcWMOQQwd2peEBBE&m=mk1COTV1sAZu82fBqU9P70ZPQXi-d6NrV1-5QYTPHOo&s=fzrCGjX9wfPKGZuwd7u4KJ4_AyzK0ZQtU9_PRyCam3U&e=) appears to be syntactically incorrect

...

...

"'Äpfel' will not be found..." contains text "Apfel" diacritics

sensitive

...

...

In the BaseX GUI the keyword diacritics is underlined in red and the

following error is reported

...

...

Unexpected end of query: 'diacritic sens...'.

This happens in version 8.6.4 and also the latest (9.0.2).

Thanks, Ron

Ron Katriel, Ph.D. | Principal Data Scientist | Medidata Solutions

350 Hudson Street, 7th Floor, New York, NY 10014

rkatriel@mdsol.com | direct: +1 201 337 3622 | mobile: +1 201 675

5598 | main: +1 212 918 1800

...

...

Hi Christian,

Thanks for the advise. The BaseX engine is phenomenal so I realized quickly that the problem was performing a naive cross product.

Since this query is run only once a month (to serialize XML to CSV) and applied to new data (DB) each time, a BaseX map will likely be the most straightforward solution (I used the same idea for another project with good results).

I will not be able to implement and test this for another couple of weeks but will summarize my findings to the group as soon as possible.

Best, Ron

On Aug 4, 2018, at 6:00 AM, Christian Grün christian.gruen@gmail.com wrote:

Hi Ron,

...

I believe the slow execution may be due to a combinatorial issue: the cross product of 280,000 clinical trials and ~10,000 drugs in DrugBank (not counting synonyms).

Yes, this sounds like a pretty expensive operation. Having maps (XQuery, Java) will be much faster indeed.

As Gerrit suggested, and if you will run your query more than once, it would definitely be another interesting option to build an auxiliary, custom "index database" that allows you to do exact searches (this database may still have references to your original data sets). Since version 9 of BaseX, volatile hash maps will be created for looped string comparisons. See the following example:

let $values1 := (1 to 500000) ! string() let $values2 := (500001 to 1000000) ! string() return $values1[. = $values2]

Algorithmically, 500'000 * 500'000 string comparisons will need to be performed, resulting in a total of 250 billion operations (and no results). The runtime is much faster as you might expect (and, as far as I can judge, much faster than in any other XQuery processor).

Best, Christian

Maps that reference nodes are pointers, rather than copies. It sounds like you could map every drug name to every "interesting" XML node that contains it using grouping during map creation and then just iterate on the keys to process the nodes.

On Sun, Sep 2, 2018 at 4:52 PM Ron Katriel rkatriel@mdsol.com wrote:

Hi Christian,

As promised here is a summary of my experimentation. I replaced the expensive join with a map lookup and the program finished in 4 minutes vs. 1 hour using a naive loop over the two databases (the original 6 hours reported were due to overly aggressive virus scanning software, which I turned off for this benchmarking).

The downside of not using “contains text” inside the double loop (due to its slowness) is that I had to tokenize the CT.gov interventions and remove stopwords prior to looking them up in the DrugBank map. This is a subpar solution as some drugs are missed (looking up all the possible word combinations would be expensive).

It would be nice if there was a way to combine the matching flexibility of the “contains text” construct (with its myriad of options) and the efficiency of a map lookup but that may require a finite-state automaton such as the Aho–Corasick algorithm. If you are aware of any existing solutions I would appreciate your sharing them.

Thanks, Ron

On August 4, 2018 at 8:47:49 PM, Ron Katriel (rkatriel@mdsol.com) wrote:

Hi Christian,

Thanks for the advise. The BaseX engine is phenomenal so I realized quickly that the problem was performing a naive cross product.

Since this query is run only once a month (to serialize XML to CSV) and applied to new data (DB) each time, a BaseX map will likely be the most straightforward solution (I used the same idea for another project with good results).

I will not be able to implement and test this for another couple of weeks but will summarize my findings to the group as soon as possible.

Best, Ron

...

On Aug 4, 2018, at 6:00 AM, Christian Grün christian.gruen@gmail.com

wrote:

...

Hi Ron,

...

I believe the slow execution may be due to a combinatorial issue: the

cross product of 280,000 clinical trials and ~10,000 drugs in DrugBank (not counting synonyms).

...

Yes, this sounds like a pretty expensive operation. Having maps (XQuery, Java) will be much faster indeed.

As Gerrit suggested, and if you will run your query more than once, it would definitely be another interesting option to build an auxiliary, custom "index database" that allows you to do exact searches (this database may still have references to your original data sets). Since version 9 of BaseX, volatile hash maps will be created for looped string comparisons. See the following example:

let $values1 := (1 to 500000) ! string() let $values2 := (500001 to 1000000) ! string() return $values1[. = $values2]

Algorithmically, 500'000 * 500'000 string comparisons will need to be performed, resulting in a total of 250 billion operations (and no results). The runtime is much faster as you might expect (and, as far as I can judge, much faster than in any other XQuery processor).

Best, Christian

Let's suppose you've got a map like: (and that by just typing this into the email I haven't left in any really horrible typos!)

let $drugInfo as map(xs:string,element()) := map:merge( for $element in collection('newDrugInfo')/descendant::infoElement let $name as xs:string := (: whatever you do to extract the official drug name from the update data :) return map:entry($name,$element))

then in the other docbase you've got: let $updatePlaces as map(xs:string,element()+) := map:merge( for $place in collection('updating-this-one')/descendant::couldBeInteresting let $drugName as xs:string := (:whatever you're doing now to match the drug name; there's an assumption you expect to find only one :) where exists($drugName) (: because you might not have one! :) group by $drugName (:baseX will magically make $place be a sequence of all the $place values with this drug name, effectively a sequence of pointers to those element nodes) return map:entry($drugName,$place) )

So now you can: for $drug in map:keys($drugInfo) (: we're iterating through the official list :) let $needsUpdate as element()+ := $updatePlaces($drug) for $place in $needsUpdate (: iterate through our sequence of pointers :) (: do whatever you're doing to insert the information in $drugInfo($drug) :)

It looks like the same old n-squared inner-loop/outer-loop update process, but I have found that it doesn't perform like that. I am almost never updating the docbase so whatever magic is involved may go away when you do that, but I've found this "map both sides" pattern to be very useful when merging data.

-- Graydon

On Sun, Sep 2, 2018 at 9:25 PM Ron Katriel rkatriel@mdsol.com wrote:

Hi Graydon,

Thanks for the suggestion. Could you provide sample code to help with this? If needed I can share the relevant BaseX snippet.

Best, Ron

On Sep 2, 2018, at 9:16 PM, Graydon Saunders graydonish@gmail.com wrote:

Maps that reference nodes are pointers, rather than copies. It sounds like you could map every drug name to every "interesting" XML node that contains it using grouping during map creation and then just iterate on the keys to process the nodes.

On Sun, Sep 2, 2018 at 4:52 PM Ron Katriel rkatriel@mdsol.com wrote:

...

Hi Christian,

As promised here is a summary of my experimentation. I replaced the expensive join with a map lookup and the program finished in 4 minutes vs. 1 hour using a naive loop over the two databases (the original 6 hours reported were due to overly aggressive virus scanning software, which I turned off for this benchmarking).

The downside of not using “contains text” inside the double loop (due to its slowness) is that I had to tokenize the CT.gov https://urldefense.proofpoint.com/v2/url?u=http-3A__CT.gov&d=DwMFaQ&c=fi2D4-9xMzmjyjREwHYlAw&r=44jDQvzmnB_-ovfO6Iusj0ItciJrcWMOQQwd2peEBBE&m=BBeNdMyieZ0Pe47vEYrOWalNS4uDt0n0tRpqmay-0Ug&s=ThHquw8-y4wRz3ejr0mTEm-ImDJHZ_DKVjr8_laQAps&e= interventions and remove stopwords prior to looking them up in the DrugBank map. This is a subpar solution as some drugs are missed (looking up all the possible word combinations would be expensive).

It would be nice if there was a way to combine the matching flexibility of the “contains text” construct (with its myriad of options) and the efficiency of a map lookup but that may require a finite-state automaton such as the Aho–Corasick algorithm. If you are aware of any existing solutions I would appreciate your sharing them.

Thanks, Ron

On August 4, 2018 at 8:47:49 PM, Ron Katriel (rkatriel@mdsol.com) wrote:

Hi Christian,

Thanks for the advise. The BaseX engine is phenomenal so I realized quickly that the problem was performing a naive cross product.

Since this query is run only once a month (to serialize XML to CSV) and applied to new data (DB) each time, a BaseX map will likely be the most straightforward solution (I used the same idea for another project with good results).

I will not be able to implement and test this for another couple of weeks but will summarize my findings to the group as soon as possible.

Best, Ron

...

On Aug 4, 2018, at 6:00 AM, Christian Grün christian.gruen@gmail.com

wrote:

...

Hi Ron,

...

I believe the slow execution may be due to a combinatorial issue: the

cross product of 280,000 clinical trials and ~10,000 drugs in DrugBank (not counting synonyms).

...

Yes, this sounds like a pretty expensive operation. Having maps (XQuery, Java) will be much faster indeed.

As Gerrit suggested, and if you will run your query more than once, it would definitely be another interesting option to build an auxiliary, custom "index database" that allows you to do exact searches (this database may still have references to your original data sets). Since version 9 of BaseX, volatile hash maps will be created for looped string comparisons. See the following example:

let $values1 := (1 to 500000) ! string() let $values2 := (500001 to 1000000) ! string() return $values1[. = $values2]

Algorithmically, 500'000 * 500'000 string comparisons will need to be performed, resulting in a total of 250 billion operations (and no results). The runtime is much faster as you might expect (and, as far as I can judge, much faster than in any other XQuery processor).

Best, Christian