inscriptis - HTML to text converter.
Project description
A python based HTML to text conversion library, command line client and Web service with support for nested tables, a subset of CSS and optional support for providing an annotated output.
Inscriptis is particularly well suited for applications that require high-performance, high-quality (i.e., layout-aware) text representations of HTML content, and will aid knowledge extraction and data science tasks conducted upon Web data.
Please take a look at the Rendering document for a demonstration of inscriptis’ conversion quality.
A Java port of inscriptis 1.x has been published by x28.
This document provides a short introduction to Inscriptis.
The full documentation is built automatically and published on Read the Docs.
If you are interested in a more general overview on the topic of text extraction from HTML, this blog post on different HTML to text conversion approaches, and criteria for selecting them might be interesting to you.
Statement of need - why inscriptis?
Inscriptis provides a layout-aware conversion of HTML that more closely resembles the rendering obtained from standard Web browsers and, therefore, better preserves the spatial arrangement of text elements.
Conversion quality becomes a factor once you need to move beyond simple HTML snippets. Non-specialized approaches and less sophisticated libraries do not correctly interpret HTML semantics and, therefore, fail to properly convert constructs such as itemizations, enumerations, and tables.
Beautiful Soup’s get_text() function, for example, converts the following HTML enumeration to the string firstsecond.
<ul> <li>first</li> <li>second</li> <ul>
Inscriptis, in contrast, not only returns the correct output
* first * second
but also supports much more complex constructs such as nested tables and also interprets a subset of HTML (e.g., align, valign) and CSS (e.g., display, white-space, margin-top, vertical-align, etc.) attributes that determine the text alignment. Any time the spatial alignment of text is relevant (e.g., for many knowledge extraction tasks, the computation of word embeddings and language models, and sentiment analysis) an accurate HTML to text conversion is essential.
Inscriptis supports annotation rules, i.e., user-provided mappings that allow for annotating the extracted text based on structural and semantic information encoded in HTML tags and attributes used for controlling structure and layout in the original HTML document. These rules might be used to
provide downstream knowledge extraction components with additional information that may be leveraged to improve their respective performance.
assist manual document annotation processes (e.g., for qualitative analysis or gold standard creation). Inscriptis supports multiple export formats such as XML, annotated HTML and the JSONL format that is used by the open source annotation tool doccano.
enabling the use of Inscriptis for tasks such as content extraction (i.e., extract task-specific relevant content from a Web page) which rely on information on the HTML document’s structure.
Installation
At the command line:
$ pip install inscriptis
Or, if you don’t have pip installed:
$ easy_install inscriptis
Python library
Embedding inscriptis into your code is easy, as outlined below:
import urllib.request
from inscriptis import get_text
url = "https://www.fhgr.ch"
html = urllib.request.urlopen(url).read().decode('utf-8')
text = get_text(html)
print(text)
Standalone command line client
The command line client converts HTML files or text retrieved from Web pages to the corresponding text representation.
Command line parameters
The inscript command line client supports the following parameters:
usage: inscript [-h] [-o OUTPUT] [-e ENCODING] [-i] [-d] [-l] [-a] [-r ANNOTATION_RULES] [-p POSTPROCESSOR] [--indentation INDENTATION] [--table-cell-separator TABLE_CELL_SEPARATOR] [-v] [input] Convert the given HTML document to text. positional arguments: input Html input either from a file or a URL (default:stdin). optional arguments: -h, --help show this help message and exit -o OUTPUT, --output OUTPUT Output file (default:stdout). -e ENCODING, --encoding ENCODING Input encoding to use (default:utf-8 for files; detected server encoding for Web URLs). -i, --display-image-captions Display image captions (default:false). -d, --deduplicate-image-captions Deduplicate image captions (default:false). -l, --display-link-targets Display link targets (default:false). -a, --display-anchor-urls Display anchor URLs (default:false). -r ANNOTATION_RULES, --annotation-rules ANNOTATION_RULES Path to an optional JSON file containing rules for annotating the retrieved text. -p POSTPROCESSOR, --postprocessor POSTPROCESSOR Optional component for postprocessing the result (html, surface, xml). --indentation INDENTATION How to handle indentation (extended or strict; default: extended). --table-cell-separator TABLE_CELL_SEPARATOR Separator to use between table cells (default: three spaces). -v, --version display version information
HTML to text conversion
convert the given page to text and output the result to the screen:
$ inscript https://www.fhgr.ch
convert the file to text and save the output to fhgr.txt:
$ inscript fhgr.html -o fhgr.txt
convert the file using strict indentation (i.e., minimize indentation and extra spaces) and save the output to fhgr-layout-optimized.txt:
$ inscript --indentation strict fhgr.html -o fhgr-layout-optimized.txt
convert HTML provided via stdin and save the output to output.txt:
$ echo "<body><p>Make it so!</p></body>" | inscript -o output.txt
HTML to annotated text conversion
convert and annotate HTML from a Web page using the provided annotation rules.
Download the example annotation-profile.json and save it to your working directory:
$ inscript https://www.fhgr.ch -r annotation-profile.json
The annotation rules are specified in annotation-profile.json:
{
"h1": ["heading", "h1"],
"h2": ["heading", "h2"],
"b": ["emphasis"],
"div#class=toc": ["table-of-contents"],
"#class=FactBox": ["fact-box"],
"#cite": ["citation"]
}
The dictionary maps an HTML tag and/or attribute to the annotations inscriptis should provide for them. In the example above, for instance, the tag h1 yields the annotations heading and h1, a div tag with a class that contains the value toc results in the annotation table-of-contents, and all tags with a cite attribute are annotated with citation.
Given these annotation rules the HTML file
<h1>Chur</h1>
<b>Chur</b> is the capital and largest town of the Swiss canton of the
Grisons and lies in the Grisonian Rhine Valley.
yields the following JSONL output
{"text": "Chur\n\nChur is the capital and largest town of the Swiss canton
of the Grisons and lies in the Grisonian Rhine Valley.",
"label": [[0, 4, "heading"], [0, 4, "h1"], [6, 10, "emphasis"]]}
The provided list of labels contains all annotated text elements with their start index, end index and the assigned label.
Annotation postprocessors
Annotation postprocessors enable the post processing of annotations to formats that are suitable for your particular application. Post processors can be specified with the -p or --postprocessor command line argument:
$ inscript https://www.fhgr.ch \ -r ./annotation/examples/annotation-profile.json \ -p surface
Output:
{"text": " Chur\n\n Chur is the capital and largest town of the Swiss
canton of the Grisons and lies in the Grisonian Rhine Valley.",
"label": [[0, 6, "heading"], [8, 14, "emphasis"]],
"tag": "<heading>Chur</heading>\n\n<emphasis>Chur</emphasis> is the
capital and largest town of the Swiss canton of the Grisons and
lies in the Grisonian Rhine Valley."}
Currently, inscriptis supports the following postprocessors:
surface: returns a list of mapping between the annotation’s surface form and its label:
[ ['heading', 'Chur'], ['emphasis': 'Chur'] ]
xml: returns an additional annotated text version:
<?xml version="1.0" encoding="UTF-8" ?> <heading>Chur</heading> <emphasis>Chur</emphasis> is the capital and largest town of the Swiss canton of the Grisons and lies in the Grisonian Rhine Valley.
html: creates an HTML file which contains the converted text and highlights all annotations as outlined below:
Web Service
A FastAPI-based Web Service that uses Inscriptis for translating HTML pages to plain text.
Run the Web Service on your host system
Install the optional feature web-service for inscriptis:
$ pip install inscriptis[web-service]
Start the Inscriptis Web service with the following command:
$ uvicorn inscriptis.service.web:app --port 5000 --host 127.0.0.1
Run the Web Service with Docker
The docker definition can be found here:
$ docker pull ghcr.io/weblyzard/inscriptis:latest $ docker run -n inscriptis ghcr.io/weblyzard/inscriptis:latest
Run as Kubernetes Deployment
The helm chart for deployment on a kubernetes cluster is located in the inscriptis-helm repository.
Use the Web Service
The Web services receives the HTML file in the request body and returns the corresponding text. The file’s encoding needs to be specified in the Content-Type header (UTF-8 in the example below):
$ curl -X POST -H "Content-Type: text/html; encoding=UTF8" \ --data-binary @test.html http://localhost:5000/get_text
The service also supports a version call:
$ curl http://localhost:5000/version
Example annotation profiles
The following section provides a number of example annotation profiles illustrating the use of Inscriptis’ annotation support. The examples present the used annotation rules and an image that highlights a snippet with the annotated text on the converted web page, which has been created using the HTML postprocessor as outlined in Section annotation postprocessors.
Wikipedia tables and table metadata
The following annotation rules extract tables from Wikipedia pages, and annotate table headings that are typically used to indicate column or row headings.
{
"table": ["table"],
"th": ["tableheading"],
"caption": ["caption"]
}
The figure below outlines an example table from Wikipedia that has been annotated using these rules.
References to entities, missing entities and citations from Wikipedia
This profile extracts references to Wikipedia entities, missing entities and citations. Please note that the profile isn’t perfect, since it also annotates [ edit ] links.
{
"a#title": ["entity"],
"a#class=new": ["missing"],
"class=reference": ["citation"]
}
The figure shows entities and citations that have been identified on a Wikipedia page using these rules.
Posts and post metadata from the XDA developer forum
The annotation rules below, extract posts with metadata on the post’s time, user and the user’s job title from the XDA developer forum.
{
"article#class=message-body": ["article"],
"li#class=u-concealed": ["time"],
"#itemprop=name": ["user-name"],
"#itemprop=jobTitle": ["user-title"]
}
The figure illustrates the annotated metadata on posts from the XDA developer forum.
Code and metadata from Stackoverflow pages
The rules below extracts code and metadata on users and comments from Stackoverflow pages.
{
"code": ["code"],
"#itemprop=dateCreated": ["creation-date"],
"#class=user-details": ["user"],
"#class=reputation-score": ["reputation"],
"#class=comment-date": ["comment-date"],
"#class=comment-copy": ["comment-comment"]
}
Applying these rules to a Stackoverflow page on text extraction from HTML yields the following snippet:
Advanced topics
Annotated text
Inscriptis can provide annotations alongside the extracted text which allows downstream components to draw upon semantics that have only been available in the original HTML file.
The extracted text and annotations can be exported in different formats, including the popular JSONL format which is used by doccano.
Example output:
{"text": "Chur\n\nChur is the capital and largest town of the Swiss canton
of the Grisons and lies in the Grisonian Rhine Valley.",
"label": [[0, 4, "heading"], [0, 4, "h1"], [6, 10, "emphasis"]]}
The output above is produced, if inscriptis is run with the following annotation rules:
{
"h1": ["heading", "h1"],
"b": ["emphasis"],
}
The code below demonstrates how inscriptis’ annotation capabilities can be used within a program:
import urllib.request
from inscriptis import get_annotated_text
from inscriptis.model.config import ParserConfig
url = "https://www.fhgr.ch"
html = urllib.request.urlopen(url).read().decode('utf-8')
rules = {'h1': ['heading', 'h1'],
'h2': ['heading', 'h2'],
'b': ['emphasis'],
'table': ['table']
}
output = get_annotated_text(html, ParserConfig(annotation_rules=rules)
print("Text:", output['text'])
print("Annotations:", output['label'])
Fine tuning
The following options are available for fine tuning inscriptis’ HTML rendering:
More rigorous indentation: call inscriptis.get_text() with the parameter indentation='extended' to also use indentation for tags such as <div> and <span> that do not provide indentation in their standard definition. This strategy is the default in inscript and many other tools such as Lynx. If you do not want extended indentation you can use the parameter indentation='standard' instead.
Overwriting the default CSS definition: inscriptis uses CSS definitions that are maintained in inscriptis.css.CSS for rendering HTML tags. You can override these definitions (and therefore change the rendering) as outlined below:
from lxml.html import fromstring
from inscriptis.css_profiles import CSS_PROFILES, HtmlElement
from inscriptis.html_properties import Display
from inscriptis.model.config import ParserConfig
# create a custom CSS based on the default style sheet and change the
# rendering of `div` and `span` elements
css = CSS_PROFILES['strict'].copy()
css['div'] = HtmlElement(display=Display.block, padding=2)
css['span'] = HtmlElement(prefix=' ', suffix=' ')
html_tree = fromstring(html)
# create a parser using a custom css
config = ParserConfig(css=css)
parser = Inscriptis(html_tree, config)
text = parser.get_text()
Custom HTML tag handling
If the fine-tuning options discussed above are not sufficient, you may even override Inscriptis’ handling of start and end tags as outlined below:
from inscriptis import ParserConfig
from inscriptis.html_engine import Inscriptis
from inscriptis.model.tag import CustomHtmlTagHandlerMapping
my_mapping = CustomHtmlTagHandlerMapping(
start_tag_mapping={'a': my_handle_start_a},
end_tag_mapping={'a': my_handle_end_a}
)
inscriptis = Inscriptis(html_tree,
ParserConfig(custom_html_tag_handler_mapping=my_mapping))
text = inscriptis.get_text()
In the example the standard HTML handlers for the a tag are overwritten with custom versions (i.e., my_handle_start_a and my_handle_end_a). You may define custom handlers for any tag, regardless of whether it already exists in the standard mapping.
Please refer to custom-html-handling.py for a working example. The standard HTML tag handlers can be found in the inscriptis.model.tag package.
Optimizing memory consumption
Inscriptis uses the Python lxml library which prefers to reuse memory rather than release it to the operating system. This behavior might lead to an increased memory consumption, if you use inscriptis within a Web service that parses very complex HTML pages.
The following code mitigates this problem on Unix systems by manually forcing lxml to release the allocated memory:
import ctypes
def trim_memory() -> int:
libc = ctypes.CDLL("libc.so.6")
return libc.malloc_trim(0)
Examples
Strict indentation handling
The following example demonstrates modifying ParserConfig for strict indentation handling.
from inscriptis import get_text
from inscriptis.css_profiles import CSS_PROFILES
from inscriptis.model.config import ParserConfig
config = ParserConfig(css=CSS_PROFILES['strict'].copy())
text = get_text('fi<span>r</span>st', config)
print(text)
Ignore elements during parsing
Overwriting the default CSS profile also allows changing the rendering of selected elements. The snippet below, for example, removes forms from the parsed text by setting the definition of the form tag to Display.none.
from inscriptis import get_text
from inscriptis.css_profiles import CSS_PROFILES, HtmlElement
from inscriptis.html_properties import Display
from inscriptis.model.config import ParserConfig
# create a custom CSS based on the default style sheet and change the
# rendering of `div` and `span` elements
css = CSS_PROFILES['strict'].copy()
css['form'] = HtmlElement(display=Display.none)
# create a parser configuration using a custom css
html = """First line.
<form>
User data
<label for="name">Name:</label><br>
<input type="text" id="name" name="name"><br>
<label for="pass">Password:</label><br>
<input type="hidden" id="pass" name="pass">
</form>"""
config = ParserConfig(css=css)
text = get_text(html, config)
print(text)
Citation
There is a Journal of Open Source Software paper you can cite for Inscriptis:
@article{Weichselbraun2021,
doi = {10.21105/joss.03557},
url = {https://doi.org/10.21105/joss.03557},
year = {2021},
publisher = {The Open Journal},
volume = {6},
number = {66},
pages = {3557},
author = {Albert Weichselbraun},
title = {Inscriptis - A Python-based HTML to text conversion library optimized for knowledge extraction from the Web},
journal = {Journal of Open Source Software}
}
Changelog
A full list of changes can be found in the release notes.
Project details
Release history Release notifications | RSS feed
Download files
Download the file for your platform. If you're not sure which to choose, learn more about installing packages.
Source Distribution
Built Distribution
Hashes for inscriptis-2.5.0-py3-none-any.whl
Algorithm | Hash digest | |
---|---|---|
SHA256 | 28ac4eed83bc470833eada0139d07672a74b4f7eb4dbab69c93ca1e903de2ea0 |
|
MD5 | 7cb6f5ecbb6cb8254e4a7605c8c31020 |
|
BLAKE2b-256 | 36efb5459a0192ad8a943be67a4821e5234be9b567e7b8a34658c93a460c51d6 |