nlabel is a library for generating, storing and retrieving tagging information and embedding vectors from various nlp libraries through a unified interface.

Overview

Logo

nlabel is currently alpha software and in an early stage of development.

nlabel is a library for generating, storing and retrieving tagging information and embedding vectors from various nlp libraries through a unified interface.

nlabel is also a system to collate results from various taggers and keep track of used models and configurations.

Apart from its standard persistence through sqlite and json files, nlabel's binary arriba format combines especially low storage requirements with high performance (see benchmarks below).

Through arriba, nlabel is thus especially suitable for

  • inspecting many features on few documents
  • inspecting few features on many documents

To support external tool chains, nlabel supports exporting to REFI-QDA.

Quick Start

Processing text works occurs in two steps. First, a NLP instance is built from an existing NLP pipeline:

from nlabel import NLP

import spacy

nlp = NLP(spacy.load("en_core_web_sm"), renames={
    'pos': 'upos',
    'tag': 'xpos'
}, require_gpu=False)

In the example, above nlp now contains a pipeline based on spacy's en_core_web_sm model. We instruct nlp to generate embedding vectors via vectors, and to rename two tags, namely pos to upos and tag to xpos.

In the next step, we run the pipeline and look at its output:

doc = nlp(
    "If you're going to San Francisco,"
    "be sure to wear some flowers in your hair.")

for sent in doc.sentences:
    for token in doc.tokens:
        print(token.text, token.upos, token.vector)

You can ask a doc which tags it carries, by calling doc.tags. In the example above, this would give:

['dep', 'ent_iob', 'lemma', 'morph', 'sentence', 'token', 'upos', 'xpos']

In the following sections, some of internal concepts will be explained. To get directly to code that will generate archives for document collections, skip to Importing a CSV to a local archive.

Tags and Labels

nlabel handles everything as tags, even it is has no label. That means that nlabel regards tokens and sentences as as tags with labels. Tags can both be iterated but also asked for labels. Tags can also be regarded as containers that contain other tags. The following examples illustrate the concepts:

for ent in doc.ents:
    print(ent.label, ent.text)

outputs

GPE San Francisco,

while

for ent in doc.ents:
    for token in ent.tokens:
        print(ent.label, token.text, token.xpos)

outputs

GPE San PROPN
GPE Francisco PROPN

NLP engines

To plug in a different nlp engine, set nlp differently:

import stanza
nlp = NLP(stanza.Pipeline('en'))

Since we renamed tag and pos, in the spacy example above, this would work without additional work.

At the moment nlabel has implementations for spacy, stanza, flair and deeppavlov. You can also write your own nlp data generators (based on nlabel.nlp.Tagger).

While NLP usually auto-detects the type of NLP parser you provide it, there are specialized constructors (NLP.spacy, NLP.flair, etc.) that cover some border cases.

Saving and Loading Documents

Documents can be saved to disk:

doc.save("path/to/file")

By default, this will generate a json-based format that should be easy to parse, even if you do decide to not use nlabel after this point - see bahia json documentation.

Of course, you can also use nlabel to load its own documents:

from nlabel import Document

with Document.open("path/to/file") as doc:
    for sent in doc.sentences:
        for token in sent.tokens:
            print(token.text, token.upos, token.vector)

Working with Archives

To store data from multiple taggers and texts, the approach from the last section would generate lots of separate files. nlabel offers a much better alternative through Archives.

There will be more detailed info on archives later on, for now, here is a quick run-through of how to use them.

A first example

This creates (or opens an existing) archive using the carenero engine (details later on), and adds a newly parsed document to it.

with open_archive("/path/to/archive", engine="carenero") as archive:
    doc = nlp(text)
    archive.add(doc)

Opening the archive later would allow us to retrieve all documents:

with open_archive("/path/to/archive", "r") as archive:
    for doc in archive.iter():
        print(doc.text)

Archives know some more things like the number of documents - use len(archive) - or information about its taggers (see next section).

Multiple Taggers

Things get interesting when using more than one tagger, e.g.:

with open_archive("/path/to/archive", engine="carenero") as archive:
    archive.add(nlp1(text))  # e.g. spacy
    archive.add(nlp2(text))  # e.g. stanza

In such an archive, calling archive.iter() will produce an error:

there are 2 taggers with conflicting tag names in this archive,
please use a selector

The reason for this error message is that spacy's and stanza's tag names clash, and nlabel would not know how to deciper doc.tokens to map either to spacy's or stanza's token data.

To resolve this issue, we can specify which tagger to use in iter.

To do this, we can first ask the archive for the taggers it knows by calling archive.taggers. Each tagger carries a unique signature that identifies it. For example, print(archive.taggers[0]) might the following signature:

env:
  machine: arm64
  platform: macOS-12.1-arm64-arm-64bit
  runtime:
    nlabel: 0.0.1.dev0
    python: 3.9.7
library:
  name: spacy
  version: 3.2.1
model:
  lang: en
  name: core_web_sm
  version: 3.2.0
renames:
  pos: upos
  tag: xpos
type: nlp
vectors:
  token:
    type: native

To iterate over documents getting tag data from this tagger, we can use archive.iter(archive.taggers[0]).

More commonly, we want to select a tagger based on its attributes, not on its index in an archive. To do this, we can use a MongoDB style query syntax:

spacy_tagger = archive.taggers[{
    'library': {
        'name': 'spacy'
    }
}]

This will return the tagger, that carries the name 'spacy' in the 'library' section of its signature. If there are no or multiple such taggers, we will get a KeyError.

As shorthand for the query above, you can also use:

spacy_tagger = archive.taggers[{
    'library.name': 'spacy'
}]

Mixing and Bridging Taggers

What happens if we want not exactly one tagger, but the output from multiple taggers.

Archive.iter() also allows to specify single tags and even rename them.

Using spacy_tagger from the last section and a new stanza_tagger:

for doc in archive.iter( spacy_tagger.sentence, spacy_tagger.xpos, stanza_tagger.xpos.to('st_xpos'))):

With these docs, we now can access spacy's sentence and xpos tags, but also stanza's xpos tag, which we rename to st_xpos to avoid a name clash with spacy's `xpos' tag:

    for token in doc.tokens:  # spacy tokens
        print(token.xpos)  # spacy xpos
        print(token.st_xpos)  # stanza xpos

Note that this only works, if stanza's tokenization for a token exactly matches that of spacy.

The Design of nlabel and Inherent Quirks

nlabel does not differentiate between tags and structuring entities such as sentences and tokens. All of them are the same concept to nlabel: labeled spans, that can be containers to other spans.

What can look like a bug at times, is a very conscious design decision: nlabel is completely agnostic to tags in terms of knowing only a single concept that it applies to everything.

Due to this design, there are various formulations in the API that are perfectly valid but rather confusing.

Obviously, it is desirable to write code that avoids these valid but quirky formulations.

Anything is a span with a label

The code below will look for a tag called "pos" that is perfectly aligned with the current token. If such a tag exists, nlabel considers it to be the "token's pos tag", and will return this tag's label.

for token in doc.tokens:
    print(token.pos)

Here is a quirky twist on the code above:

for token in doc.tokens:
    print(token.sentence)

This is allowed. The code will do the same thing as above: first it looks for a tag called "sentence" that is perfectly aligned with the current token. If such a tag exists, its label is returned.

Since the "sentence" tags provided by nlp libraries carry no labels, and "sentence" tags are not aligned to "token" tags, this will fail at step one or two, and therefore just return an empty label. Still, it is valid in terms of nlabel's concepts.

Using the "label" attribute

for ent in sentence.ents:
    print(ent.label)

The following code does exactly the same thing (avoid using it):

for ent in sentence.ents:
    print(ent.ent)

Label Types

There are four label types in nlabel:

description notes
labels all labels constisting of value and score
label first label only ignores ensuing labels
strs string list of label values ignores scores
str first label value as string ignores score and ensuing labels

strs and labels are suitable for getting output from taggers that return multiple labels.

The default type is str. The exception to this rule are morphology tags (e.g. spacy's morph and stanza's feats, which default to strs).

To specify label types, use the .to(label_type=x) method on tags, when specifying them to Archive.iter or Group.view.

Groups and Views

Groups are an underlying building block of nlabel. You might not encounter them directly.

A group contains data from multiple taggers for one shared text. If you need to collect data for multiple texts, use archives.

Documents can be combined into Groups, which will then contain information from multiple taggers:

from nlabel import Group

group = Group.join([doc1, doc2])

Groups have a view method that works similar to the iter method available in Archives.

Computing Embeddings

The following code uses a spacy model to generate token vectors from spacy's native vector attribute:

nlp = NLP.spacy(
    spacy_model,
    vectors={'token': nlabel.embeddings.native})

Spacy's vector attribute is usually filled via spacy's own Tok2Vec and Transformer components or external extensions such as spacy-sentence-bert.

Alternatively, the following code constructs a model that computes transformer embeddings for tokens via flair:

nlp = NLP.flair(
    vectors={'token': nlabel.embeddings.huggingface(
        "dbmdz/bert-base-german-cased", layers="-1, -2")},
    from_spacy=spacy_model)

from_spacy indicates that sentence splitter and tokenizer should be taken from the provided spacy model.

Archives

Engines

nlabel comes with three different persistence engines:

  • carenero is for collecting data, esp. in a batch setting - by supporting restartability and transaction safety, and enabling export of full data or sub sets of it into bahia or arriba.
  • bahia is suitable for archival purposes, as it is just a thin wrapper around a zip of human-readable json files; it is not the ideal format for exports.
  • arriba is a binary format optimized for read performance, it is suitable for data analysis; it is not suitable for exports.

Storage Size

The following graph shows data from a real-world dataset, consisting of 18861 texts (125.3 MB text data), tagged with 4 taggers and a total of 31 tags (no embedding data). Y axis shows size in GB (note logarithmic scale). REFI-QDA is roughly 100 times the size of arriba.

storage size requirements for different engines

Random Access Speeds

The exact speed of arriba depends on the task and data, but but often arriba performs 10 to 100 times faster than bahia and carenero on real-world projects. From the same data set as earlier (when extracting all POS tags from one of 4 taggers over 2000 documents):

access times for different engines

The carenero/ALL benchmarks shows the time when accessing all tags from all taggers through carenero.

More Engine Details

These engines support storing both tagging data and embedding vectors. In the ordering above, they go from slower to faster.

carenero bahia arriba
data collection + - -
exporting + - -
read speeds - - +
suitable for archival - + -

(*) bahia supports writes, but does not avoid adding duplicates or support proper restartability in batch settings, i.e. it is not suited to incremental updates.

Additional Examples

Importing a CSV to a local archive

Create a carenero archive from a CSV:

from nlabel.importers import CSV

import spacy

csv = CSV(
    "/path/to/some.csv",
    keys=['zeitung_id', 'text_type_id', 'filename'],
    text='text')
csv.importer(spacy.load("en_core_web_sm")).to_local_archive()

This will create an archive located in the same folder as the CSV. The code above is restartable, i.e. it is okay to interrupt and continue later - it will not add duplicate entries.

Once the archive has been created, one can either use it directly, e.g. iterating its documents:

from nlabel import open_archive

with open_archive("some/archive.nlabel", mode="r") as archive:
    for doc in archive.iter(some_selector):
        for x in doc.tokens:
            print(x.text, x.xpos, x.vector)

Or, one can save the archive to different formats for faster traversal:

archive.save("demo2", engine="bahia")
archive.save("demo3", engine="arriba")

The open_archive call from above works with all archive types.

Note that the iter call on archives takes an optional view description that allows picking/renaming tags as described earlier.

Exporting to a remote archive

For larger jobs, it is often useful to separate computation and storage, and to allow multiple computation processes (both often applies to GPU cluster environments). Since carenero's sqlite is bad at handling concurrent writes, the solution is starting a dedicated web service that handles the writing on a dedicated machine.

On machine A, start an archive server (it will write a carenero archive to the given path):

python -m nlabel.importers.server /path/to/archive.nlabel --password your_pwd

On machine B, you can start one or multiple importers writing to that remote archive. Modifying the example from the local archive:

from nlabel import RemoteArchive

remote_archive = RemoteArchive("http://localhost:8000", ("user", "your_pwd"))
csv.importer(spacy.load("en_core_web_sm")).to_remote_archive(
    remote_archive, batch_size=8)

Exporting REFI-QDA

The following code exports ent tags to a REFI-QDA project.

from nlabel import NLP

import spacy
nlp = NLP(spacy.load("en_core_web_lg"))
text = 'some longer text...'
doc = nlp(text)

doc.save_to_qda(
    "/path/to/your.qdp", {
        'tagger': {
        },
        'tags': {
            'ent'
        }
    })

A save_to_qda method is also part of cantenero and bahia archives.

Owner
Bernhard Liebl
Bernhard Liebl
🧪 Cutting-edge experimental spaCy components and features

spacy-experimental: Cutting-edge experimental spaCy components and features This package includes experimental components and features for spaCy v3.x,

Explosion 65 Dec 30, 2022
PUA Programming Language written in Python.

pua-lang PUA Programming Language written in Python. Installation git clone https://github.com/zhaoyang97/pua-lang.git cd pua-lang pip install . Try

zy 4 Feb 19, 2022
GPT-Code-Clippy (GPT-CC) is an open source version of GitHub Copilot, a language model

GPT-Code-Clippy (GPT-CC) is an open source version of GitHub Copilot, a language model -- based on GPT-3, called GPT-Codex -- that is fine-tuned on publicly available code from GitHub.

Nathan Cooper 2.3k Jan 01, 2023
This repository contains the code, models and datasets discussed in our paper "Few-Shot Question Answering by Pretraining Span Selection"

Splinter This repository contains the code, models and datasets discussed in our paper "Few-Shot Question Answering by Pretraining Span Selection", to

Ori Ram 88 Dec 31, 2022
PeCo: Perceptual Codebook for BERT Pre-training of Vision Transformers

PeCo: Perceptual Codebook for BERT Pre-training of Vision Transformers

Microsoft 105 Jan 08, 2022
Learning General Purpose Distributed Sentence Representations via Large Scale Multi-task Learning

GenSen Learning General Purpose Distributed Sentence Representations via Large Scale Multi-task Learning Sandeep Subramanian, Adam Trischler, Yoshua B

Maluuba Inc. 309 Oct 19, 2022
code for modular summarization work published in ACL2021 by Krishna et al

This repository contains the code for running modular summarization pipelines as described in the publication Krishna K, Khosla K, Bigham J, Lipton ZC

Approximately Correct Machine Intelligence (ACMI) Lab 21 Nov 24, 2022
Awesome Treasure of Transformers Models Collection

💁 Awesome Treasure of Transformers Models for Natural Language processing contains papers, videos, blogs, official repo along with colab Notebooks. 🛫☑️

Ashish Patel 577 Jan 07, 2023
Technique for Order of Preference by Similarity to Ideal Solution (TOPSIS)

TOPSIS implementation in Python Technique for Order of Preference by Similarity to Ideal Solution (TOPSIS) CHING-LAI Hwang and Yoon introduced TOPSIS

Hamed Baziyad 8 Dec 10, 2022
中文医疗信息处理基准CBLUE: A Chinese Biomedical LanguageUnderstanding Evaluation Benchmark

English | 中文说明 CBLUE AI (Artificial Intelligence) is playing an indispensabe role in the biomedical field, helping improve medical technology. For fur

452 Dec 30, 2022
Fake news detector filters - Smart filter project allow to classify the quality of information and web pages

fake-news-detector-1.0 Lists, lists and more lists... Spam filter list, quality keyword list, stoplist list, top-domains urls list, news agencies webs

Memo Sim 1 Jan 04, 2022
PyTorch implementation of the paper: Text is no more Enough! A Benchmark for Profile-based Spoken Language Understanding

Text is no more Enough! A Benchmark for Profile-based Spoken Language Understanding This repository contains the official PyTorch implementation of th

Xiao Xu 26 Dec 14, 2022
Proquabet - Convert your prose into proquints and then you essentially have Vogon poetry

Proquabet Turn your prose into a constant stream of encrypted and meaningless-so

Milo Fultz 2 Oct 10, 2022
An Explainable Leaderboard for NLP

ExplainaBoard: An Explainable Leaderboard for NLP Introduction | Website | Download | Backend | Paper | Video | Bib Introduction ExplainaBoard is an i

NeuLab 319 Dec 20, 2022
Code for the ACL 2021 paper "Structural Guidance for Transformer Language Models"

Structural Guidance for Transformer Language Models This repository accompanies the paper, Structural Guidance for Transformer Language Models, publis

International Business Machines 10 Dec 14, 2022
Code for text augmentation method leveraging large-scale language models

HyperMix Code for our paper GPT3Mix and conducting classification experiments using GPT-3 prompt-based data augmentation. Getting Started Installing P

NAVER AI 47 Dec 20, 2022
Coreference resolution for English, German and Polish, optimised for limited training data and easily extensible for further languages

Coreferee Author: Richard Paul Hudson, msg systems ag 1. Introduction 1.1 The basic idea 1.2 Getting started 1.2.1 English 1.2.2 German 1.2.3 Polish 1

msg systems ag 169 Dec 21, 2022
The code for the Subformer, from the EMNLP 2021 Findings paper: "Subformer: Exploring Weight Sharing for Parameter Efficiency in Generative Transformers", by Machel Reid, Edison Marrese-Taylor, and Yutaka Matsuo

Subformer This repository contains the code for the Subformer. To help overcome this we propose the Subformer, allowing us to retain performance while

Machel Reid 10 Dec 27, 2022
Shirt Bot is a discord bot which uses GPT-3 to generate text

SHIRT BOT · Shirt Bot is a discord bot which uses GPT-3 to generate text. Made by Cyclcrclicly#3420 (474183744685604865) on Discord. Support Server EX

31 Oct 31, 2022
YACLC - Yet Another Chinese Learner Corpus

汉语学习者文本多维标注数据集YACLC V1.0 中文 | English 汉语学习者文本多维标注数据集(Yet Another Chinese Learner

BLCU-ICALL 47 Dec 15, 2022