Text Analytics On Company Reviews In Jobstreet

Data Collection

The data using in this work are company reviews posted by companies’ employees in JobStreet website. At the initial stage of work, three well-known companies in Penang were selected, namely Company A, Company B and Company C. A total of 132 reviews posted from 01 Jan 2017 to 10 Dec 2018 were collected from JobStreet website ( JobStreet.com Malaysia, 2018) and the review distribution of the companies is shown in Figure 02 .

See Full Size >

The data size might be insufficient to provide a comprehensive information about the company but it should be sufficient to provide basic significant insight. The company reviews collected from JobStreet website were transformed into Excel table for further processing. The data collection and transformation process are summarised in Figure 03 . As shown in Figure 03 , there are two parts of reviews, i.e., “the good things” and “the challenges” which represented the positive and negative reviews about the company respectively. Thus, separate analysis was required for these two parts of reviews.

See Full Size >

Data exploration

After the data collection, exploration and inspection were performed to ensure the validity of the data collected. There were 132 rows and 6 columns in the collected data. The 132 rows are the reviews for the three companies. The descriptions for the six columns are shown below (see Table 01 ). Data exploration on data type for each column was as performed to make sure no inappropriate format is used, as well as no missing value. All collected data are stored in Excel table. After that, the dataset is ready for pre-processing stage.

Text Pre-processing

In this stage, data were categorized as “the_good_things” and “the_challenges” which represented the positive and negative insights of company, respectively. Since the data is unstructured data, therefore data cleaning is required. Based on Loper and Bird ( 2002), there are six steps taken in the text pre-processing process: first, tokenize the text data by using unigram tokenization; second, remove the punctuations in text data; third, remove extra whitespace between each word for standardization; fourth, convert the text to lowercase for standardization purpose. However, exception for certain words such as “US” and “IT” were applied to avoid the deviation of the word meanings. The list of exception words was designed in scalable format in Python script where new exception word can be added easily in future; fifth, remove numeric text which is not meaningful in representing the insight about company; and finally remove stop words that have no significant meaning such as “the” and “of”. The stop word list is a predefined list in Python Natural Language Toolkit (NLTK) library. In terms of tokenization, this work uses the unigram tokenization as initial evaluation (to check the feasibility to apply text analytics in job searching process). A simple testing was also performed in Online NGram Analyzer using bi-gram or more tokenization to check if there are any significant outputs. The findings will be further discussed in Text Mining section.

Text Normalization

After the text pre-processing step, the cleaned text data require additional step for standardization. The wording or format standardization step is called text normalization. In general, there are two types of normalization, which are stemming and lemmatization. Stemming is a process to convert the word to its stem which is a good method to standardize different wording with same stem. However, stemming was not used because it would produce invalid wording which could bringing difficulty to understand the outputs of analytics (word clouds). Lemmatization is another method to standardize the text. The output of lemmatization is valid word which overcome the issue in stemming. It might not as good as stemming in standardizing the word to stem word, but it is good enough in basic wording standardization such as plural convert to singular. Thus, lemmatization is applied to normalize the text data in this work.

Text Mining

In order to identify the significant insights from the company reviews, basic text analytic technique which is term frequency or word count has been applied. An assumption is that the frequent term in the reviews would be more likely representing the insight about the company. There are two types of graphics used to represent the frequent words: word cloud and histogram. The word cloud provides quick glance insight about the company for job seekers. Meanwhile the histogram visualize the quantitative information about word frequency. This graphical output was only be used as a reference in the development stage, not to deliver as final output for job seekers.

As the objective of work is to provide a quick insight about positive and negative of each company, thus, there would be different word clouds to represent positive and negative reviews. Since there are three companies were selected for in this study, a total of six word clouds will be generated to represent the positive and negative comments about the three companies.

In the analysis of preliminary graphics developed from normalized text data, words with less significant meanings were found. These less meaningful words such as “also” and “may” were categorized as stop words, which was then removed during the text pre-processing stage. Besides, it was also found that the preliminary word cloud was rather crowded as showed in Figure 04 , which caused distraction to identify the meaningful keyword.

See Full Size >

Thus, term reduction was applied to reduce the term with frequency of 2 or below as it is least likely to represent the company insight. However, only 24 reviews were available for Company C which is half the size of the reviews available for Company A and Company B. After removed the term with frequency ≤ 2 for Company C, it was only left less than 20 terms. Thus, the term reduction for frequency ≤ 1 was applied for the reviews of Company C. In overall, 83.7% of terms were reduced as shown in Table 02 .

By removing the identified less meaningful word and apply term reduction, the new word clouds have been developed and illustrated in Table 03 .

Table 04 below shows some of the positive (good things) and negative (challenges) insights about the company obtained from the six Word Clouds.

As mentioned in Text Pre-processing section, a quick analysis is conducted using Online NGram Analyzer ( Ngram Analyzer, 2018) by applying bi-gram or more tokenization to explore if there is any additional information. There are some frequent and significant ngram term found as showed in Table 05 . The table shows that bi-gram or tri-gram tokenization can produce more specific and new insight such as career growth in Company A, helpful manager and the challenge of converting from contract worker to a permanent position in Company B.

However, Online NGram Analyzer is not a good platform to achieve the objective of this work due to the manual step needed to input the text, plus there is no text normalization feature (e.g. stemming and lemmatization). Online NGram Analyzer was just used as a quick check to assess whether different n-gram provide additional insights about the company in this work. Thus, expanding the unigram word cloud developed by Python to cover bi-gram or more tokenization is a potential future enhancement.

Discussion

The objective of work is to provide a platform for job seekers to gain quick insight about the company. In this work, word clouds used as platform to provide such insight to the job seekers. Overall, the word cloud (refer to Table 03 ) is able to produce accurate and informative insight as evaluated by respective employees. However, there are still potential gaps to be improved on the accuracy or to gain more insight by focusing on the factors and their potential future work (see Table 08 ).

There are some challenges faced in this work such as size of data etc. Thus, future enhancement is needed to produce more accurate and informative outputs. As a preparation for future enhancement, the Python script was designed in a user friendly and scalable structure. The stop word list and lowercase exception list is expandable from time to time for newly detected word.

Other than job seekers, this solution (word clouds) is also applicable for company human resource (HR) personal for better understand the employee perspective and take necessary action as an employee retention approach. Besides, the HR personal can also understand the benefit offered by competitor companies in analysing the employee’s needs and leaving reasons.