Data collection

Data Collection

without code snippet:

Data preprocess

• We removed code snippets, HTML tags.
• We then converted all documents to lower-case words token, and remove punctuation signs.
• And removed common English words (i.e.,stop-words), did lemmatization, keeping only noun, adj, vb, adv.
• After that, we transformed words into their root (i.e., stemming). We used the improved version of the Porter stemming for this purpose [20].
• In order to further reduce noise from data we followed the approach suggested by Yang et al., [28]. We formed the word frequency matrix out of the whole data corpus and removed the words that appeared less than 10 times in total.

LDA

1. Latent Dirichlet Allocation (LDA) modeling technique [13] was adopted for this paper.

2. We calculated topic coherence scores for a range of 2-100 topics where we fixed alpha, beta with default, and arrived at 15 as the value in the range of the highest values (i.e. most coherent topic model).

   topics	c_v
   4	0.49238079391933737
   15	0.4803587450047443
   5	0.48020645922082095
   3	0.4800883619689366
   6	0.47612252341174427
   12	0.47392385923203323
   7	0.4688837055506901
   17	0.4658745233276905
   18	0.4642817316407959
   10	0.463303793989635

3. Next, we calculated the optimal hyperparameter alpha and beta parameters with 15 topics which have highest coherence score, and finally got the best parameter value.

   alpha	beta	topics	c_v
   0.41	0.81	15	0.4953024941697182

4. We assigned a descriptive label to each of the topics. To do so, for each topic we carefully analyzed the top 10 terms and manually checked some of the highly relevant documents of the topic.

Topic #4 top 10 documents

Topic #2 top 10 documents

Topic #14 top 10 documents