Skip to main content

TEXTA Multilingual Processor (MLP)

Project description

TEXTA MLP Python package

http://pypi.texta.ee/texta-mlp/

Installation

Requirements

apt-get install python3-lxml

From PyPI

pip3 install texta-mlp

From Git

pip3 install git+https://git.texta.ee/texta/texta-mlp-python.git

Testing

python3 -m pytest -v tests

Usage

Load MLP

Supported languages: https://stanzanlp.github.io/stanzanlp/models.html

>>> from texta_mlp.mlp import MLP
>>> mlp = MLP(language_codes=["et","en","ru"])

Process & Lemmatize Estonian

>>> mlp.process("Selle eestikeelse lausega võiks midagi ehk öelda.")
{'text': {'text': 'Selle eestikeelse lausega võiks midagi ehk öelda .', 'lang': 'et', 'lemmas': 'see eestikeelne lause võima miski ehk ütlema .', 'pos_tags': 'P A S V P J V Z'}, 'texta_facts': []}
>>>
>>> mlp.lemmatize("Selle eestikeelse lausega võiks midagi ehk öelda.")
'see eestikeelne lause võima miski ehk ütlema .'

You can use the "analyzers" argument to limit the amount of data you want to be analyzed and returned, thus speeding up the process. Accepted options are: ["lemmas", "pos_tags", "transliteration", "ner", "contacts", "entity_mapper", "all"] where "all" signifies that you want to use all analyzers (takes the most time). By the default, this value is "all".

>>> mlp.process("Selle eestikeelse lausega võiks midagi ehk öelda.", analyzers=["lemmas", "postags"])

Process & Lemmatize Russian

>>> mlp.process("Лукашенко заявил о договоренности Москвы и Минска по нефти.")
{'text': {'text': 'Лукашенко заявил о договоренности Москвы и Минска по нефти .', 'lang': 'ru', 'lemmas': 'лукашенко заявить о договоренность москва и минск по нефть .', 'pos_tags': 'X X X X X X X X X X', 'transliteration': 'Lukašenko zajavil o dogovorennosti Moskvõ i Minska po nefti .'}, 'texta_facts': []}
>>>
>>> mlp.lemmatize("Лукашенко заявил о договоренности Москвы и Минска по нефти.")
'лукашенко заявить о договоренность москва и минск по нефть .

Process & Lemmatize English

>>> mlp.process("Test sencences are rather difficult to come up with.")
{'text': {'text': 'Test sencences are rather difficult to come up with .', 'lang': 'en', 'lemmas': 'Test sencence be rather difficult to come up with .', 'pos_tags': 'NN NNS VBP RB JJ TO VB RB IN .'}, 'texta_facts': []}
>>>
>>> mlp.lemmatize("Test sencences are rather difficult to come up with.")
'Test sencence be rather difficult to come up with .'

Make MLP Throw an Exception on Unknown Languages

By default, MLP will default to Estonian if language is unknown. To not do so, one must provide use_default_language_code=False when initializing MLP.

>>> mlp.process("المادة 1 يولد جميع الناس أحرارًا متساوين في الكرامة والحقوق. وقد وهبوا عقلاً وضميرًا وعليهم أن يعامل بعضهم بعضًا بروح الإخاء.")
{'text': {'text': 'المادة 1 يولد جميع الناس أحرارًا متساوين في الكرامة والحقوق . وقد وهبوا عقلاً وضميرًا وعليهم أن يعامل بعضهم بعضًا بروح الإخاء .', 'lang': 'et', 'lemmas': 'lee 1 يولد جميع الناس leele leele في leele leele . وقد وهبوا عقلاً leele lee أن يعامل بعضهم بعضًا بروح lee .', 'pos_tags': 'S N S S S S S S S S Z S S S S S S S S Y Y Y Z'}, 'texta_facts': []}
>>>
>>> mlp = MLP(language_codes=["et","en","ru"], use_default_language_code=False)
>>> mlp.process("المادة 1 يولد جميع الناس أحرارًا متساوين في الكرامة والحقوق. وقد وهبوا عقلاً وضميرًا وعليهم أن يعامل بعضهم بعضًا بروح الإخاء.")
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/home/rsirel/dev/texta-mlp-package/texta_mlp/mlp.py", line 150, in process
    document = self.generate_document(raw_text, loaded_analyzers)
  File "/home/rsirel/dev/texta-mlp-package/texta_mlp/mlp.py", line 96, in generate_document
    lang = self.detect_language(processed_text)
  File "/home/rsirel/dev/texta-mlp-package/texta_mlp/mlp.py", line 89, in detect_language
    raise LanguageNotSupported("Detected language is not supported: {}.".format(lang))
texta_mlp.exceptions.LanguageNotSupported: Detected language is not supported: ar.

Change Default Language Code

Do use some other language as default, one must provide default_language_code when initializing MLP.

>>> mlp = MLP(language_codes=["et", "en", "ru"], default_language_code="en")
>>>
>>> mlp.process("المادة 1 يولد جميع الناس أحرارًا متساوين في الكرامة والحقوق. وقد وهبوا عقلاً وضميرًا وعليهم أن يعامل بعضهم بعضًا بروح الإخاء.")
{'text': {'text': 'المادة 1 يولد جميع الناس أحرارًا متساوين في الكرامة والحقوق . وقد وهبوا عقلاً وضميرًا وعليهم أن يعامل بعضهم بعضًا بروح الإخاء .', 'lang': 'en', 'lemmas': 'المادة 1 يولد جميع الناس أحرارًا متساوين في الكرامة والحقوق . وقد وهبوا عقلاً وضميرًا وعليهم أن يعامل بعضهم بعضًا بروح الإخاء .', 'pos_tags': 'NN CD , NN NN NN NN IN NN NN . UH NN NN NN NN NN NN NN NN NN NN .'}, 'texta_facts': []}

Process Arabic (for real this time)

>>> mlp = MLP(language_codes=["et","en","ru", "ar"])
>>> mlp.process("المادة 1 يولد جميع الناس أحرارًا متساوين في الكرامة والحقوق. وقد وهبوا عقلاً وضميرًا وعليهم أن يعامل بعضهم بعضًا بروح الإخاء.")
{'text': {'text': 'المادة 1 يولد جميع الناس أحرارًا متساوين في الكرامة والحقوق . وقد وهبوا عقلاً وضميرًا وعليهم أن يعامل بعضهم بعضا بروح الإخاء .', 'lang': 'ar', 'lemmas': 'مَادَّة 1 وَلَّد جَمِيع إِنسَان حَرَر مُتَسَاوِي فِي كَرَامَة والحقوق . وَقَد وَ عَقَل وضميراً وعليهم أَنَّ يعامل بعضهم بَعض بروح إِخَاء .', 'pos_tags': 'N------S1D Q--------- VIIA-3MS-- N------S4R N------P2D N------P4I A-----MP4I P--------- N------S2D U--------- G--------- U--------- VP-A-3MP-- N------S4I A-----MS4I U--------- C--------- VISA-3MS-- U--------- N------S4I U--------- N------S2D G---------', 'transliteration': "AlmAdp 1 ywld jmyE AlnAs >HrArFA mtsAwyn fy AlkrAmp wAlHqwq . wqd whbwA EqlAF wDmyrFA wElyhm >n yEAml bEDhm bEDA brwH Al<xA' ."}, 'texta_facts': []}
>>>
>>> mlp.lemmatize("المادة 1 يولد جميع الناس أحرارًا متساوين في الكرامة والحقوق. وقد وهبوا عقلاً وضميرًا وعليهم أن يعامل بعضهم بعضا بروح الإخاء.")
'مَادَّة 1 وَلَّد جَمِيع إِنسَان حَرَر مُتَسَاوِي فِي كَرَامَة والحقوق . وَقَد وَ عَقَل وضميراً وعليهم أَنَّ يعامل بعضهم بَعض بروح إِخَاء .'

Load MLP with Custom Resource Path

>>> mlp = MLP(language_codes=["et","en","ru"], resource_dir="/home/kalevipoeg/mlp_resources/")

Different phone parsers

Texta MLP has three different phone parsers:

  • 'phone_strict' - is used by default. It parses only those numbers that are verified by the phonenumbers library. It verifies all correct numbers if they have an area code before it. Otherwise (without an area code) it verifies only Estonian ("EE") and Russian ("RU") phone numbers. This is because in this example "Maksekorraldusele märkida viitenumber 2800049900 ning selgitus ...", the "2800049900" is a valid number in Great Britain ("GB"), but not with "EE" or "RU".

  • 'phone_high_precision' which output is mainly phonenumbers extracted by regex, but the regex excludes complicated versions.

  • 'phone_high_recall' was originally done for emails and it gets most of the phone numbers (includes complicated versions), but also outputs a lot of noisy data. This parser is also used by default in concatenating close entities (read below). This means that while concatenating, only "PHONE_high_recall" fact is considered and other parsers' results are not included in concatenating (avoids overlaping). The other parsers' results won't get lost and are still added in texta_facts. Just not under the fact "BOUNDED".

You can choose the parsers like so:

>>> mlp.process(analyzers=["lemmas", "phone_high_precision"], raw_text= "My phone number is 12 34 56 77.")

Concatenate close entities

Let`s test MLP() and Concatenator() on the following three letters. Letter 1:

Dear all, 

Let`s not forget that I intend to concure the whole of Persian Empire!

Best wishes,
Alexander Great
aleksandersuur356eKr@mail.ee
phone: 76883266

Letter 2:

От: Terry Pratchett < tpratchett@gmail.com >
Кому: Joe Abercrombie < jabercrombie@gmail.com >
Название: Разъяснение

Дорогой Joe,

Как вы? Надеюсь, у тебя все хорошо. Последний месяц я писал свой новый роман, 
который обещал представить в начале лета. Я тоже немного почитал и обожаю твою 
новую книгу!

Я просто хотел уточнить, что Alexander Great жил в Македонии.

Лучший,
Terry

Letter 3:

Dear Terry!

Terry Pratchett already created Discworld. This name is taken. Other than that I found 
the piece fascanating and see great potential in you! I strongly encourage you to take 
action in publishing your works. Btw, if you would like to show your works to Pratchett 
as well, he`s interested. I talked about you to him. His email is tpratchett@gmail.com. 
Feel free to write him!

Joe


From: Terry Berry < bigfan@gmail.com >
To: Joe Abercrombie < jabercrombie@gmail.com >
Title: Question

Hi Joe,

I finally finished my draft and I`m sending it to you. The hardest part 
was creating new places. What do you think of the names of the places I created?

Terry Berry
 

Let`s read all those letters into a list called "mailbox". We will process the letters as discribed above and save them into a jsonlines file.

from texta_mlp.mlp import MLP
mlp = MLP(language_codes=["et","en","ru"])
processed_letters = []
for letter in mailbox:
    processed_letters += [mlp.process(letter)]
   
import jsonlines
with jsonlines.open("letters.jsonl", mode="w") as writer:
    writer.write_all(processed_letters)

MLP() already creates a fact BOUNDED which bounds the closest entities within the letter together. In order to sort out the info in whole mailbox we have to concatenate the BOUNDED facts. It means creating a database of personal info gotten from different letters. For that we use the Concatenator(), which input is processed letters.

from texta_mlp.concatenator import Concatenator

cn = Concatenator()
cn.load_bounded_from_jsonl(path = "letters.jsonl")
#cn.load_bounded_fron_jsonl() uses default path "mlpanalyzed.jsonl"

Then we will concatenate the BOUNDED facts. Be aware that with large mailboxes it might take 2 hours!

cn.concatenate()

We can check the length of the database lists and the content with functions:

  • cn._just_pers_infos() (type "close_persons", persons appearing close in letter(s)),
  • cn._bounded() (the original unconcatenated bounded),
  • cn._unsure_infos() (type "unsure_whose_entities", enities that have >=2 candidate persons, not sure to whom it belongs),
  • cn._no_personas_infos() (type "no_per_close_entities", entities appearing close in letter(s) without persons nearby),
  • cn._persona_infos() (type "person_info", the real deal, entities with its person).

All of that can be saved to .jsonl file.

cn.save_to_jsonlines(path="concatenated_bounds_from_mailbox.jsonl")
#cn.save_to_jsonlines() uses default path "concatenated_bounds.jsonl"

Output of "concatenated_bounds_from_mailbox.jsonl":

{"type": "person_info", "PER": "Alexander Great", "LOC": ["Македония", "Persian Empire"], "EMAIL": ["aleksandersuur356eKr@mail.ee"], "PHONE": ["76883266"]}
{"type": "person_info", "PER": "Joe Abercrombie", "EMAIL": ["jabercrombie@gmail.com"]}
{"type": "person_info", "PER": "Terry Berry", "EMAIL": ["bigfan@gmail.com"]}
{"type": "person_info", "PER": "Terry Pratchett", "EMAIL": ["tpratchett@gmail.com"]}

Dealing with Elasticsearch

We can also use Elasticsearch with Concatenator(). Here`s a snippet for getting from Elasticsearch and processing documents already processed by MLP() and then uploading them to a new index.

from texta_mlp.concatenator import Concatenator
cn = Concatenator()
cn.load_bounded_from_elastic(es_url= 'http://localhost:8888', index_name = "mlp_processed_mails")
cn.concatenate()
cn.save_to_elasticsearch(index_name = 'http://localhost:8888', es_url = "mails_concatenated_bounded")

Using just cn.load_bounded_from_elastic() uses default settings:

cn.load_bounded_from_elasticsearch(es_url= 'http://elastic-dev.texta.ee:9200', index_name = "mlp_processed_mails")

Using just cn.save_to_elasticsearch() uses default settings:

cn.save_to_elasticsearch(index_name = 'http://elastic-dev.texta.ee:9200', es_url = "concatenated_BOUNDED")

Project details


Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

texta-mlp-1.5.2.tar.gz (44.9 kB view hashes)

Uploaded Source

Supported by

AWS AWS Cloud computing and Security Sponsor Datadog Datadog Monitoring Fastly Fastly CDN Google Google Download Analytics Microsoft Microsoft PSF Sponsor Pingdom Pingdom Monitoring Sentry Sentry Error logging StatusPage StatusPage Status page