unitok 2.2.2.0

Unified Tokenizer

Unified Tokenizer

Introduction

When dealing with textual information used in some models (e.g. Bert), the first step is the tokenization. Transformers provides BertTokenizer to split words (multi-lingual) with WordPiece algorithm. However, in some cases, some texts may be entities which is not required to split, although this entity contains many sub-words; some are the arrays of entities, joined by some characters (e.g. |, ,).

Different cases require different tokenizer. So it comes Unified Tokenizer (or UniTok). You can either customize your tokenizer, or use our pre-defined tokenizers.

Installation

pip install UniTok

Usage

Here we use the first 10 lines of the training set of MINDlarge as an example (see news-sample.tsv file). Assume the data path is /home/ubuntu/news-sample.tsv.

Data Declaration (more info see MIND GitHub)

Each line in the file is the information about one piece of news.

It has 7 columns, which are divided by the tab symbol:

• News ID
• Category
• SubCategory
• Title
• Abstract
• URL
• Title Entities (entities contained in the title of this news)
• Abstract Entities (entites contained in the abstract of this news)

We only use its first 5 columns for demonstration.

Imports

import pandas as pd


from UniTok import UniTok, Column
from UniTok.tok import IdTok, EntTok, BertTok, SingT, ListT

Read data

df = pd.read_csv(
    filepath_or_buffer='/home/ubuntu/news-sample.tsv',
    sep='\t',
    names=['nid', 'cat', 'subCat', 'title', 'abs', 'url', 'titEnt', 'absEnt'],
    usecols=['nid', 'cat', 'subCat', 'title', 'abs'],
)

Initialize Tokenizers

id_tok = IdTok(name='news')  # for news id
cat_tok = EntTok(name='cat')  # for category & subcategory
txt_tok = BertTok(name='english', vocab_dir='bert-base-uncased')  # for title & abstract
cat_tok.vocab.reserve(100)  # first 100 tokens are reserved for some special usage in the downstream model, if any, and please be reminded that the first token is always PAD

Construct Unified Tokenizer

SingleTokenizer means it only omits one token id, while ListTokenizer generates a sequence of ids.

ut = UniTok()
ut.add_col(Column(
    name='nid',
    tokenizer=id_tok.as_sing(),
)).add_col(Column(
    name='cat',
    tokenizer=cat_tok.as_sing()
)).add_col(Column(
    name='subCat',
    tokenizer=cat_tok.as_sing(),
)).add_col(Column(
    name='title',
    tokenizer=txt_tok.as_list(),
)).add_col(Column(
    name='abs',
    tokenizer=txt_tok.as_list(),
)).read_file(df)

Here we leave the max_length of the output of the ListTokenizer behind.

Analyse Data

ut.analyse()

It shows the distribution of the length of each column (if using ListTokenizer). It will help us determine the max_length of the tokens for each column.

[ COLUMNS ]
[ COL: nid ]
[NOT ListTokenizer]

[ COL: cat ]
[NOT ListTokenizer]

[ COL: subCat ]
[NOT ListTokenizer]

[ COL: title ]
[ MIN: 6 ]
[ MAX: 16 ]
[ AVG: 12 ]
[ X-INT: 1 ]
[ Y-INT: 0 ]
       |   
       |   
       |   
       |   
       || |
       || |
       || |
| |  | || |
| |  | || |
| |  | || |
-----------

[ COL: abs ]
100%|██████████| 10/10 [00:00<00:00, 119156.36it/s]
100%|██████████| 10/10 [00:00<00:00, 166440.63it/s]
100%|██████████| 10/10 [00:00<00:00, 164482.51it/s]
100%|██████████| 10/10 [00:00<00:00, 2172.09it/s]
100%|██████████| 10/10 [00:00<00:00, 1552.30it/s]
[ MIN: 0 ]
[ MAX: 46 ]
[ AVG: 21 ]
[ X-INT: 1 ]
[ Y-INT: 0 ]
|                                              
|                                              
|                                              
|                                              
|                                              
|               | | ||    ||               |  |
|               | | ||    ||               |  |
|               | | ||    ||               |  |
|               | | ||    ||               |  |
|               | | ||    ||               |  |
-----------------------------------------------

[ VOCABS ]
[ VOC: news with  10 tokens ]
[ COL: nid ]

[ VOC: cat with  112 tokens ]
[ COL: cat, subCat ]

[ VOC: english with  30522 tokens ]
[ COL: title, abs ]

ReConstruct Unified Tokenizer

ut = UniTok()
ut.add_col(Column(
    name='nid',
    tokenizer=id_tok.as_sing(),
)).add_col(Column(
    name='cat',
    tokenizer=cat_tok.as_sing()
)).add_col(Column(
    name='subCat',
    tokenizer=cat_tok.as_sing(),
)).add_col(Column(
    name='title',
    tokenizer=txt_tok.as_list(max_length=10),
)).add_col(Column(
    name='abs',
    tokenizer=txt_tok.as_list(max_length=30),
)).read_file(df)

In this step, we set max_length of each column. If max_length is not set, we will keep the whole sequence and not truncate it.

Tokenize and Store

ut.tokenize()
ut.store_data('TokenizedData')