Update README.md
Browse files
README.md
CHANGED
|
@@ -1,3 +1,199 @@
|
|
| 1 |
---
|
| 2 |
license: cc-by-sa-3.0
|
| 3 |
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
---
|
| 2 |
license: cc-by-sa-3.0
|
| 3 |
---
|
| 4 |
+
|
| 5 |
+
# The Wikipedia Webpage 2M (WikiWeb2M) Dataset
|
| 6 |
+
|
| 7 |
+
We present the WikiWeb2M dataset consisting of over 2 million English
|
| 8 |
+
Wikipedia articles. Our released dataset includes all of the text content on
|
| 9 |
+
each page, links to the images present, and structure metadata such as which
|
| 10 |
+
section each text and image element comes from.
|
| 11 |
+
|
| 12 |
+
This dataset is a contribution from our [paper](https://arxiv.org/abs/2305.03668)
|
| 13 |
+
`A Suite of Generative Tasks for Multi-Level Multimodal Webpage Understanding`.
|
| 14 |
+
|
| 15 |
+
The dataset is stored as gzipped TFRecord files which can be downloaded here or on our [GitHub repository](https://github.com/google-research-datasets/wit/blob/main/wikiweb2m.md).
|
| 16 |
+
|
| 17 |
+
## WikiWeb2M Statistics
|
| 18 |
+
|
| 19 |
+
WikiWeb2M is the first multimodal open source dataset to include all page
|
| 20 |
+
content in a unified format. Here we provide aggregate information about the
|
| 21 |
+
WikiWeb2M dataset as well as the number of samples available with each of the
|
| 22 |
+
fine-tuning tasks we design from it.
|
| 23 |
+
|
| 24 |
+
| Number of | Train | Validation | Test |
|
| 25 |
+
| ---- | ---- | ---- | ---- |
|
| 26 |
+
| Pages | 1,803,225 | 100,475 | 100,833 |
|
| 27 |
+
| Sections | 10,519,294 | 585,651 | 588,552 |
|
| 28 |
+
| Unique Images | 3,867,277 | 284,975 | 286,390 |
|
| 29 |
+
| Total Images | 5,340,708 | 299,057 | 300,666 |
|
| 30 |
+
|
| 31 |
+
Our data processing and filtering choices for each fine-tuning task are
|
| 32 |
+
described in the paper.
|
| 33 |
+
|
| 34 |
+
| Downstream Task Samples | Train | Validation | Test |
|
| 35 |
+
| ---- | ---- | ---- | ---- |
|
| 36 |
+
| Page Description Generation | 1,435,263 | 80,103 | 80,339 |
|
| 37 |
+
| Section Summarization | 3,082,031 | 172,984 | 173,591 |
|
| 38 |
+
| Contextual Image Captioning | 2,222,814 | 124,703 | 124,188 |
|
| 39 |
+
|
| 40 |
+
|
| 41 |
+
## Data and Task Examples
|
| 42 |
+
|
| 43 |
+
Here we illustrate how a single webpage can be processed into the three tasks we
|
| 44 |
+
study: page description generation, section summarization, and contextual image
|
| 45 |
+
captioning. The paper includes multiple Wikipedia article examples.
|
| 46 |
+
|
| 47 |
+

|
| 48 |
+
|
| 49 |
+
|
| 50 |
+
|
| 51 |
+
## Usage
|
| 52 |
+
|
| 53 |
+
### TFRecord Features
|
| 54 |
+
|
| 55 |
+
Here we provide the names of the fields included in the dataset, their
|
| 56 |
+
tensorflow Sequence Example type, their data type, and a brief description.
|
| 57 |
+
|
| 58 |
+
|
| 59 |
+
| Feature | Sequence Example Type | DType | Description |
|
| 60 |
+
| ---- | ---- | ---- | ---- |
|
| 61 |
+
| `split` | Context | string | Dataset split this page contributes to (e.g., train, val, or test) |
|
| 62 |
+
| `page_url` | Context | string | Wikipeda page URL |
|
| 63 |
+
| `page_title` | Context | string | Wikipedia page title, title of the article |
|
| 64 |
+
| `raw_page_description` | Context | string | Wikipedia page description, which is typically the same or very similar to the content of the first (root) section of the article |
|
| 65 |
+
| `clean_page_description` | Context | string | `raw_page_description` but with newline and tab characters removed; this provides the exact target text for our page description generation task |
|
| 66 |
+
| `page_contains_images` | Context | int64 | Whether the Wikipedia page has images after our cleaning and processing steps |
|
| 67 |
+
| `page_content_sections_without_table_list` | Context | int64 | Number of content sections with text or images that do not contain a list or table. This field can be used to reproduce data filtering for page description generation |
|
| 68 |
+
| `is_page_description_sample` | Context | int64 | Whether a page is used as a sample for the page description fine-tuning task |
|
| 69 |
+
| `section_title` | Sequence | string | Titles of each section on the Wikipedia page, in order |
|
| 70 |
+
| `section_index` | Sequence | int64 | Index of each section on the Wikipedia page, in order |
|
| 71 |
+
| `section_depth` | Sequence | int64 | Depth of each section on the Wikipedia page, in order |
|
| 72 |
+
| `section_heading_level` | Sequence | int64 | Heading level of each section on the Wikipedia page, in order |
|
| 73 |
+
| `section_subsection_index` | Sequence | int64 | Subsection indices, grouped by section in order |
|
| 74 |
+
| `section_parent_index` | Sequence | int64 | The parent section index of each section, in order |
|
| 75 |
+
| `section_text` | Sequence | string | The body text of each section, in order |
|
| 76 |
+
| `is_section_summarization_sample` | Sequence | int64 | Whether a section is used as a sample for the section summarization fine-tuning task |
|
| 77 |
+
| `section_raw_1st_sentence` | Sequence | string | The processed out first sentence of each section, in order |
|
| 78 |
+
| `section_clean_1st_sentence` | Sequence | string | The same as `section_raw_1st_sentence` but with newline and tab characters removed. This provides the exact target text for our section summarization task |
|
| 79 |
+
| `section_rest_sentence` | Sequence | string | The processed out sentences following the first sentence of each section, in order |
|
| 80 |
+
| `section_contains_table_or_list` | Sequence | int64 | Whether section content contains a table or list; this field is needed to be able to reproduce sample filtering for section summarization |
|
| 81 |
+
| `section_contains_images` | Sequence | int64 | Whether each section has images after our cleaning and processing steps, in order |
|
| 82 |
+
| `is_image_caption_sample` | Sequence | int64 | Whether an image is used as a sample for the image captioning fine-tuning task |
|
| 83 |
+
| `section_image_url` | Sequence | string | Image URLs, grouped by section in order |
|
| 84 |
+
| `section_image_mime_type` | Sequence | string | Image mime type, grouped by section in order |
|
| 85 |
+
| `section_image_width` | Sequence | int64 | Image width, grouped by section in order |
|
| 86 |
+
| `section_image_height` | Sequence | int64 | Image height, grouped by section in order |
|
| 87 |
+
| `section_image_in_wit` | Sequence | int64 | Whether an image was originally contained in the WIT dataset, grouped by section in order |
|
| 88 |
+
| `section_image_raw_attr_desc` | Sequence | string | Image attribution description, grouped by section in order |
|
| 89 |
+
| `section_image_clean_attr_desc` | Sequence | string | The English only processed portions of the attribution description |
|
| 90 |
+
| `section_image_raw_ref_desc` | Sequence | string | Image reference description, grouped by section in order |
|
| 91 |
+
| `section_image_clean_ref_desc` | Sequence | string | The same as `section_image_raw_ref_desc` but with newline and tab characters removed; this provides the exact target text for our image captioning task |
|
| 92 |
+
| `section_image_alt_text` | Sequence | string | Image alt-text, grouped by section in order |
|
| 93 |
+
| `section_image_captions` | Sequence | string | Comma separated concatenated text from alt-text, attribution, and reference descriptions; this is how captions are formatted as input text when used |
|
| 94 |
+
|
| 95 |
+
|
| 96 |
+
### Loading the Data
|
| 97 |
+
|
| 98 |
+
Here we provide a small code snippet for how to load the TFRecord files. First,
|
| 99 |
+
load any necessary packages.
|
| 100 |
+
|
| 101 |
+
```python
|
| 102 |
+
import numpy as np
|
| 103 |
+
import glob
|
| 104 |
+
import tensorflow.compat.v1 as tf
|
| 105 |
+
from collections import defaultdict
|
| 106 |
+
```
|
| 107 |
+
|
| 108 |
+
Next, define a data parser class.
|
| 109 |
+
```python
|
| 110 |
+
class DataParser():
|
| 111 |
+
def __init__(self,
|
| 112 |
+
filepath: str = 'wikiweb2m-*',
|
| 113 |
+
path: str):
|
| 114 |
+
self.filepath = filepath
|
| 115 |
+
self.path = path
|
| 116 |
+
self.data = defaultdict(list)
|
| 117 |
+
|
| 118 |
+
def parse_data(self):
|
| 119 |
+
context_feature_description = {
|
| 120 |
+
'split': tf.io.FixedLenFeature([], dtype=tf.string),
|
| 121 |
+
'page_title': tf.io.FixedLenFeature([], dtype=tf.string),
|
| 122 |
+
'page_url': tf.io.FixedLenFeature([], dtype=tf.string),
|
| 123 |
+
'clean_page_description': tf.io.FixedLenFeature([], dtype=tf.string),
|
| 124 |
+
'raw_page_description': tf.io.FixedLenFeature([], dtype=tf.string),
|
| 125 |
+
'is_page_description_sample': tf.io.FixedLenFeature([], dtype=tf.int64),
|
| 126 |
+
'page_contains_images': tf.io.FixedLenFeature([], dtype=tf.int64),
|
| 127 |
+
'page_content_sections_without_table_list': tf.io.FixedLenFeature([] , dtype=tf.int64)
|
| 128 |
+
}
|
| 129 |
+
|
| 130 |
+
sequence_feature_description = {
|
| 131 |
+
'is_section_summarization_sample': tf.io.VarLenFeature(dtype=tf.int64),
|
| 132 |
+
'section_title': tf.io.VarLenFeature(dtype=tf.string),
|
| 133 |
+
'section_index': tf.io.VarLenFeature(dtype=tf.int64),
|
| 134 |
+
'section_depth': tf.io.VarLenFeature(dtype=tf.int64),
|
| 135 |
+
'section_heading_level': tf.io.VarLenFeature(dtype=tf.int64),
|
| 136 |
+
'section_subsection_index': tf.io.VarLenFeature(dtype=tf.int64),
|
| 137 |
+
'section_parent_index': tf.io.VarLenFeature(dtype=tf.int64),
|
| 138 |
+
'section_text': tf.io.VarLenFeature(dtype=tf.string),
|
| 139 |
+
'section_clean_1st_sentence': tf.io.VarLenFeature(dtype=tf.string),
|
| 140 |
+
'section_raw_1st_sentence': tf.io.VarLenFeature(dtype=tf.string),
|
| 141 |
+
'section_rest_sentence': tf.io.VarLenFeature(dtype=tf.string),
|
| 142 |
+
'is_image_caption_sample': tf.io.VarLenFeature(dtype=tf.int64),
|
| 143 |
+
'section_image_url': tf.io.VarLenFeature(dtype=tf.string),
|
| 144 |
+
'section_image_mime_type': tf.io.VarLenFeature(dtype=tf.string),
|
| 145 |
+
'section_image_width': tf.io.VarLenFeature(dtype=tf.int64),
|
| 146 |
+
'section_image_height': tf.io.VarLenFeature(dtype=tf.int64),
|
| 147 |
+
'section_image_in_wit': tf.io.VarLenFeature(dtype=tf.int64),
|
| 148 |
+
'section_contains_table_or_list': tf.io.VarLenFeature(dtype=tf.int64),
|
| 149 |
+
'section_image_captions': tf.io.VarLenFeature(dtype=tf.string),
|
| 150 |
+
'section_image_alt_text': tf.io.VarLenFeature(dtype=tf.string),
|
| 151 |
+
'section_image_raw_attr_desc': tf.io.VarLenFeature(dtype=tf.string),
|
| 152 |
+
'section_image_clean_attr_desc': tf.io.VarLenFeature(dtype=tf.string),
|
| 153 |
+
'section_image_raw_ref_desc': tf.io.VarLenFeature(dtype=tf.string),
|
| 154 |
+
'section_image_clean_ref_desc': tf.io.VarLenFeature(dtype=tf.string),
|
| 155 |
+
'section_contains_images': tf.io.VarLenFeature(dtype=tf.int64)
|
| 156 |
+
}
|
| 157 |
+
|
| 158 |
+
def _parse_function(example_proto):
|
| 159 |
+
return tf.io.parse_single_sequence_example(example_proto,
|
| 160 |
+
context_feature_description,
|
| 161 |
+
sequence_feature_description)
|
| 162 |
+
|
| 163 |
+
suffix = '.tfrecord*'
|
| 164 |
+
|
| 165 |
+
data_path = glob.Glob(self.path + self.filepath + suffix)
|
| 166 |
+
raw_dataset = tf.data.TFRecordDataset(data_path, compression_type='GZIP')
|
| 167 |
+
parsed_dataset = raw_dataset.map(_parse_function)
|
| 168 |
+
|
| 169 |
+
for d in parsed_dataset:
|
| 170 |
+
split = d[0]['split'].numpy().decode()
|
| 171 |
+
self.data[split].append(d)
|
| 172 |
+
```
|
| 173 |
+
|
| 174 |
+
Then you can run the following to parse the dataset.
|
| 175 |
+
```python
|
| 176 |
+
parser = DataParser()
|
| 177 |
+
parser.parse_data()
|
| 178 |
+
print((len(parser.data['train']), len(parser.data['val']), len(parser.data['test'])))
|
| 179 |
+
```
|
| 180 |
+
### Models
|
| 181 |
+
Our full attention, transient global, and prefix global experiments were run
|
| 182 |
+
using the [LongT5](https://github.com/google-research/longt5) code base.
|
| 183 |
+
|
| 184 |
+
|
| 185 |
+
## How to Cite
|
| 186 |
+
|
| 187 |
+
If you extend or use this work, please cite the [paper](https://arxiv.org/abs/2305.03668) where it was
|
| 188 |
+
introduced:
|
| 189 |
+
|
| 190 |
+
```
|
| 191 |
+
@inproceedings{
|
| 192 |
+
burns2023wiki,
|
| 193 |
+
title={A Suite of Generative Tasks for Multi-Level Multimodal Webpage Understanding},
|
| 194 |
+
author={Andrea Burns and Krishna Srinivasan and Joshua Ainslie and Geoff Brown and Bryan A. Plummer and Kate Saenko and Jianmo Ni and Mandy Guo},
|
| 195 |
+
booktitle={The 2023 Conference on Empirical Methods in Natural Language Processing (EMNLP)},
|
| 196 |
+
year={2023},
|
| 197 |
+
url={https://openreview.net/forum?id=rwcLHjtUmn}
|
| 198 |
+
}
|
| 199 |
+
```
|