The top operating organizations understand an essential role that customer feedback plays in the business industry. These businesses then consistently listen to the feedbacks of the consumers to stay ahead in the competition. Customer feedback gives crucial insights into the workings of the product, services, and what could be done within the company's domain to make the experiences of the consumers better. Customer's opinions help the companies ensure that the final product actually shall sufice their expectations, solve their problems and meet their needs. Hence, the customer feedback is one of the most reliable and easy to get sources for tangible data that can also be used in making wise business decisions. The proposed approach provides a method to make efective use of this feedback and generate insights for the Product Team. Since it is not feasible to go through all reviews to find out what the customers are talking about, the reviews are clubbed together by Topic Modelling approach. The Business Team is presented with top keywords corresponding to each group of reviews which makes it easy for them to find out the actionable areas. The way the results are presented to the team guides them in the right direction so as to improve their products and services. A model is generated once the reviews have been labelled with topics. This is helpful to classify the new reviews which keep on coming from the customers' end. The Topic Modelling algorithm is again followed once the team has good number of new reviews which will further help in improving the model.
The existing customers’ reviews are not only helpful for the new customers to find the right product but they also serve as a means for the product teams to improve their products and services. In this era of digitization, organizations use customer reviews and other feedback information from various sources and generate insights out of those reviews. Machine Learning and Natural Language Processing both are used to process these wide varieties and a huge volume of reviews. Diferent approaches such as Topic Modeling, Text Clustering are used in Natural Language Processing for Customer Feedback Analysis.
Data Preprocessing: It is an important step to preprocess textual data before performing
Natural Language Processing tasks. NLP involves text/data processing to convert the available
data into more usable and convenient form. It helps to get rid of the redundant and irrelevant
data present in the dataset and also plays a role in maintaining the standard of the text [
Topic Modeling: Topic Modeling is used for finding diferent topics from documents (basically some form of textual data) without having any knowledge in advance.
LDA (Latent Dirichlet Allocation): This topic modeling approach makes use of each
document as a diferent set of topics and every word is considered to be drawn from those topics.
A good LDA Model involves tuning of hyperparameters such as word topic density, document
topic density etc. In order to get good quality of topics, a suitable number of topics has to be
selected which can be done by measuring the Topic Coherence, which measures the degree of
semantic similarity between the words which scored highest in the topic[
Text Classification : Text Classification is a good choice to get familiar with textual data processing. It finds a lot of interesting applications in daily life. There have been a significant amount of researches in this field. One of such research is Bert Model. BERT stands for “Bidirectional Encoder Representations from Transformers”.
The remainder of this paper is structured as follows: Section 2 provides the proposed approach followed throughout the paper. In Section 3, we present the related works. In Section 4, we present the implementation of the proposed method. In Sections 5 and 6, we discuss the results, provide a conclusion and propose recommendations for some future work.
The proposed approach basically combines two aspects of Machine Learning algorithms- Clustering and Classification.
Clustering helps to avoid the manual task of labeling the product reviews by dividing them into topics or clusters. The labels then serve as the basis of classifying the new reviews. Topic Modeling finds a theme across reviews and discovers hidden topics. It can be interpreted as creating some buckets and putting each review into these buckets. First, the reviews are split into positive or negative depending on the rating value given by the customer. Then LDA Topic Modeling is used to find themes across these two categories.
The output of Topic Modeling is visualized on a webpage that displays the top Bigrams (two words frequently occurring together) corresponding to each topic or cluster identified by the LDA Topic Modeling. This type of visualization is really helpful from the perspective of the Business or Product Team as they get a clear picture of what the customers are talking about in the reviews. The team also gets the list of actual customer reviews to read them as and when needed. It also displays an Inter-topic Distance Map which reflects the clusters formed where each cluster is represented in the form of a bubble. This is very helpful for Data Scientists for analysis of the topics or clusters formed. The webpage shows a list of the most relevant words corresponding to each topic along with their frequency in the selected topic and overall frequency.
Now, we have the clusters[
The topic modeling algorithm can again be followed after a specified time (for example, after two or three months) when the Product Team has quite a reasonable quantity of new reviews. This will in result improve the quality of new topics or clusters formed.
Many researches have been done in the text summarizations and terminology identifications [
Other approaches include Topic Modeling algorithms like Latent Dirichlet Allocation, Latent Semantic Analysis etc. which enables us to discover topics from set of documents. In Topic modeling using LDA, diferent topic groups are created. It is the role of the researcher to decide the number of groups in the final output. Since there is no prior knowledge about what is the best number of groups, we generate models with diferent numbers of groups and then analyze and compare diferent topic modeling, and then the decision is made to select the topic model which is most meaningful and sensible out of all the models generated with diferent hyperparameters.
Topic Modeling is an approach which is useful in finding out the themes across the data, hence
this is quite efective when we are dealing with customer reviews. Each review is assigned one
of the themes to which it belongs with highest proportion making it easy for the businesses to
ifgure out the dificulties being faced by the customers in regard to their products and services.
4. Implementation
4.1. NLP Preprocessing
• Contractions Expansion: Contractions are quite common in English Language. The
contractions of words are created by removing specific letters and sounds. This step
expands each and every contraction to its original form to maintain the standard of the
text[
Tokenization is also known as Text Segmentation. • Lemmatization: This is one of the most important NLP preprocessing steps. Lemmatization aims at reducing a word to its base or dictionary form, which is called as the “lemma”. It really transforms words to their true root form, instead of just chopping them. For example, the words “playing”, “plays”, “played” are mapped to “play”. It can be done with the help of Python “nltk” package and makes use of a dictionary such as “WordNet” for producing the mappings. Lemmatization plays a significant role in Natural Language Processing and Artificial Intelligence tasks. In languages other than English, lemmatization can be quite complicated.
• LDA (Latent Dirichlet Allocation): LDA (Latent Dirichlet Allocation): Topic Modeling is an approach that is used to find themes across the reviews and discover hidden topics. It is based on extracting a certain number of groups consisting of specific words from the reviews. These groups represent the topics that are useful from the perspective of the Business or Product Team to find out what the customers are talking about in the reviews. LDA (Latent Dirichlet Allocation) is one of the most popular methods of Topic Modeling. LDA takes two hyperparameters into consideration, the “alpha parameter” and the “beta parameter”. The “alpha parameter” controls the mixture of topics for any given document. If it is low, the documents will have less of a mixture of topics and if it is high, the documents will have more of a mixture of topics. The “beta parameter” controls the distribution of words per topic. If it is low, the topics will likely have fewer words. If it is high, the topics will likely have more words. Another factor that LDA takes into account is K, the number of topics or groups to form. • Topic Modeling using Nouns and Adjectives: The topics generated by LDA can be a mixture of nouns, verbs, adjectives, etc. The LDA algorithm treats all tokens equally with the same importance. When we are dealing with the reviews, removing all words except nouns and adjectives helps to improve the semantic coherence of the topics. • Bigrams Formation: Bigrams refer to two words frequently occurring together in the text.
Applying LDA Topic Modeling after taking into account bigrams (or in general, n-grams) helps to improve the quality of topic models. In Python, Gensim”s Phrases model can build and implement the bigrams, trigrams, etc.
The probabilistic topic models (such as LDA) are popular approaches for textual processing and analysis. They provide predictive and latent topic representation of the corpus. It is assumed that the latent space discovered by these models is generally meaningful and useful, and evaluating such assumptions is challenging due to its unsupervised training process. Topic Coherence is a method that can be used to evaluate the LDA topics. It is based on the concept of combining a number of measures into a framework to evaluate the coherence between topics that have been generated by the model. If a set of sentences or facts support each other, they are said to be coherent. Topic Coherence measures score of a single topic by measuring the degree of semantic similarity between high importance words in the topic. Higher the value of Topic Coherence for a model, better is the quality of topics formed by the model.
Visualizing clusters makes it convenient for the Business or Product Team to evaluate, explore
and interpret the results of Cluster Analysis[
We’ve labeled the reviews after we’ve finished clustering. Each review now has a label that corresponds to the topic number to which it belongs. To perform text token processing, the BERT employs the Transformer encoder architecture. This processing is done in the full context of all tokens before and after it. Such models are pre-trained on a large corpus of text before being fine-tuned for specific NLP tasks.
BERT is an encoder stack of transformer architecture[
Finally, after combining the two aspects of Machine Learning algorithms, Clustering and Classification, we visualize the insights to see if we can have some meaningful results from them.
The customers” reviews are of utmost importance for any firm or organization. The organizations that look into the feedback given by the customers always excel in their domain. It is not possible to go through each and every piece of customer feedback manually. Clustering the reviews is a better way to get insights from them. Topic Modeling can be used to find themes across the reviews and discover hidden topics. LDA (Latent Dirichlet Allocation) is one of the most popular methods of Topic Modeling. It is a “generative probabilistic model”. After applying the LDA model, we have the topic or cluster for each customer review to which it belongs with highest probability value. As a result, we have labelled reviews, each of which belongs to one of the topics or clusters. These clusters are visualized to present them to the Product Team in an easy to interpret and analyze form.
Once the organization has a significant number of fresh reviews, it may use the clustering technique to improve the quality of topics or clusters, since we know that the more data there is, the better the model performs. This approach is quite efective from the perspective of an organization and helps them to improve the quality of their products and services by making it easy to identify actionable areas.
Another improvement that could be made in the future is to incorporate sentences or embeddings from a model like Bert into the Topic Modeling technique. The vectors from the model and LDA can be combined with some weight or hyperparameter to improve the results.
We would like to thank our college, National Institute of Technology, Kurukshetra for giving us the platform to express ourselves. Also, we would like to thank our mentor Dr. B.B. Gupta, Asst. Professor, NIT Kurukshetra.