ISNA-M

A Small Dataset of Persian News Articles from ISNA News Agency.

This dataset includes 534 news articles published between 23:36(+3:30 GMT) 09/02/1404 Jalali Calendar(29/04/2025) to 00:48(+3:30 GMT) 10/02/1404(30/04/2025).

Raw_ISNA_Dataset.csv is collected by running ISNA_crawler.py and has 10 columns with the following description:

Content	Title	Category	Source	Journalist	Secretary	Tags	Week_Day	Date	Time	URL
The content of each news article	The title of each news article	The category of each news article	The source of each news article; whether it is from an ISNA journalist or from an external source	The journalist of each news article; if available	Secretaries of each news article	Tags of each news article	The week day of the publication of each news article	The date of the publication of each news article	The time of the publication of each news article	The URL to each news article

Clean_ISNA_Dataset.csv is obtained from Raw_ISNA_Dataset.csv by running data_cleaner.py and includes two columns:

Content	Category
Lists of each article's words after preprocessing	Categories of each article after preprocessing

Some statistics of the dataset

	Raw_ISNA_Dataset.csv	Clean_ISNA_Dataset.csv
Types (# Unique Words)	20,993	13,696
Tokens (# All Words)	224,540	167,118

Ranking of words against their frequencies in logscale(Zipf's Law)

Number of tokens against types(Heap's Law)

Top 10 most frequent words

Top 10 most frequent values of each feature

Number of unique values of each feature

ISNA_crawler.py

Warning

The html tags used in this code may not work as the ISNA website's source code may change.

Scrapes ISNA website and collects data using requests, re and BeautifulSoup libraries and stores the information as a CSV file via pandas library. First, crawls on ISNA's archive page and collect news articles' URLs. Then, obtains the desired information via crawling each article's web page.

data_cleaner.py

Processes and analyses the Raw_ISNA_Dataset.csv, obtained from running ISNA_crawler.py, and stores processed contents and categories of each article to be used by category_predictor.py using re, pandas, matplotlib and parsivar libraries.

Tags, Time, Date, Week_Day, Category and Title features are processed. Content of each article is split into words and then, stop words removal and stemming are applied. The above plots show some analyses that have been done on the data.

category_predictor.py

pandas, numpy and scikitlearn libraries are used to train a random forest classifier on the Clean_ISNA_Dataset.csv, which contains Content(predictor) and Category(label) columns. Scikitlearn's CountVectorizer is used to create Bag Of Words from contents and OneHotEncoder is applied on 83 unique label values. Data is divided into training and testing subsets with 0.4 test/train ratio. Classifiers with different n_estimators values(10, 25, 50, 100, 500 and 1000) are evaluated to find the best value.

n_estimators	***	precision	recall	f1-score
10	micro avg	0.80	0.09	0.17
10	macro avg	0.08	0.03	0.04
10	weighted avg	0.21	0.09	0.12
10	samples avg	0.09	0.09	0.09
***	***	***	***	***
25	micro avg	0.74	0.09	0.17
25	macro avg	0.05	0.03	0.03
25	weighted avg	0.18	0.09	0.11
25	samples avg	0.09	0.09	0.09
***	***	***	***	***
50	micro avg	0.78	0.07	0.12
50	macro avg	0.05	0.02	0.03
50	weighted avg	0.16	0.07	0.09
50	samples avg	0.07	0.07	0.07
***	***	***	***	***
100	micro avg	0.71	0.05	0.09
100	macro avg	0.04	0.01	0.02
100	weighted avg	0.14	0.05	0.06
100	samples avg	0.05	0.05	0.05
***	***	***	***	***
500	micro avg	0.75	0.04	0.08
500	macro avg	0.03	0.01	0.02
500	weighted avg	0.10	0.04	0.05
500	samples avg	0.04	0.04	0.04
***	***	***	***	***
1000	micro avg	0.77	0.05	0.09
1000	macro avg	0.04	0.02	0.02
1000	weighted avg	0.12	0.05	0.06
1000	samples avg	0.05	0.05	0.05

The above table is the summary of the outcome. It shows that the best value of the n_estimators is 10 and others overfit the data, which is not surprising due to the small amount of the dataset.

The poor results stem from:

1- Small size of the dataset. It only contains 167,118 words. This can be tackled by collecting or synthesising more data.

2- Sparsity of the labels. Many categories have appeared less than 5 times and may appear only in training or testing subsets after spliting the data. This can be tackled by:

• Collecting or synthesing more data from underrepresented labels.
• Ignoring labels appeared once and make sure each label's samples are present in both training and testing subsets.

3- Not robust model. Other hyperparameters should be evaluated through a systematic grid search to find the best setting. Also, various architectures should be tried to find the best solution.