First Assignment - Entity Linking
Web Data Processing System Assignment 1 - 2021 - Group 26
- Zhining Bai
- Bowen Lyu
- Tianshi Chen
- Yiming Xu
Description
This is a Python program to Entity Linking by processing WARC files. We recognize entities from web pages and link them to a Knowledge Base(Wikidata). The pipeline for this program as below:
Read WARC
- Use
pysparkto read large-scale warc files, so the program supports parallel computing. - Extract text information from HTML files by using
beautifulsoup.
Named entity recognition
- Extract entities by using
recognize_entities_bertmodel fromsparknlp.
Disambiguation and NIL
We considered the popularity of the candidate page as well as the semantic similarity between the sentence where the entity is located and the candidate description to achieve Disambiguation.
- Popularity: Calculate popularity rankings using the
Elasticsearchscoring algorithm and the number of properties of the mention from the knowledge graph. - Sentence similarity: Measure the difference between text and description using the
Levenshtein distance.
NIL: Retain results with distances < 40.
Prerequisites
Codes are run on the DAS cluster at /var/scratch/wdps2106/wdps_2126, result1 is a conda virtual environment that has been created. Below are the packages installed to run the assignment.
# if you want to use pip(pip for python3) to install the packages, use the following command(python version 3.8)
pip install pyspark==3.1.2
pip install spark-nlp==3.3.3
pip install beautifulsoup4
pip install python-Levenshtein
pip install elasticsearch
# if you want to use conda to install the packages, use the following command(recommended)
conda create -n
python=3.8
conda install pyspark
conda install bs4
conda install elasticsearch
pip install python-Levenshtein
pip install sparknlp
Run
To run the program, you can simply use the command below. The parameter Keyname is the name of page ID in WARC files such as WARC_TREC_ID. You need to declare the name of the page ID using this parameter. Be aware that the result file will be renamed as result.tsv.
sh run.sh /path/to/warc/file.warc.gz /path/to/result/ Keyname
If you use DAS cluster, you also need to add this command before running:
export OPENBLAS_NUM_THREADS=10
To check the score of the result file, use the command below.
python3 score.py /sample/annotation/file/sample.tsv /generated/result/file/result.tsv
Result
We tested our entity linking code using sample.warc.gz. Since sample_annotations.tsv only contains the entities that page_id is less than 92, our test results only output entity links with page_id <= 92. The f1 score of the sample data is 0.1122.
| Metric | Value |
|---|---|
| Gold | 500 |
| Predicted | 480 |
| Correct | 55 |
| Precision | 0.1145 |
| Recall | 0.11 |
| F1 Score | 0.1122 |
21 Aug 12, 2022
2.1k Jan 1, 2023
652 Jan 6, 2023
1.2k Dec 21, 2022
19.5k Feb 13, 2021
77.3k Jan 3, 2023
12.3k Dec 31, 2022
8.4k Dec 26, 2022
3k Jan 5, 2023
30 Dec 13, 2022
21 Dec 17, 2022
2 Oct 06, 2022
2 May 17, 2022
128 Dec 29, 2022
72 Dec 07, 2022
3 Apr 18, 2022
1k Jan 08, 2023
138 Dec 27, 2022
66 Jul 05, 2022
170 Dec 13, 2022
1 Dec 30, 2021
49 Dec 30, 2022
89 Aug 12, 2022
9.1k Jan 02, 2023
3 Oct 14, 2022
50 Mar 22, 2022
2 Nov 05, 2021
177 Dec 27, 2022
0 May 06, 2022