Using Supervised Learning

Published Date: 02 Nov 2017

Introduction

Opinions are part of most of the activities that we, the human beings conduct. From our thinking and further on to the accomplishment of every aspect of our live we are surrounded by our opinions caused by sentiments, subjective and objective thinking. Being the center of the word we evaluate different situations by using our natural tool which is the opinion. Often, when we have to make a decision and we base that according to our opinions. This is a process that occurs also in organizations depending on the situation.

The importance of opinions gave birth to a new study based on the attitude, the sentiments and emotions. This field is growing rapidly with the growing of the social media such as Facebook, Twitter, blogs and other social tools. As the social network is based on our lives and our interests, we have a lot of data information regarding feelings, preferences, options and decisions. All these data are recorded and stored in digital form which makes much easier to access and analyze.

The sentiment analysis is growing as a natural language to process information. This analysis is an area of interest for data, web and text mining. Nowadays sentiment analysis is part of business activities due to its relevant studies based on what clients think, like and want. Many business activities are based on this analysis which helps them get a better understanding about the client expectations and precious feedbacks.

The data collected from different tools in social networks, blogs and other domains is usually unstructured and in the first impact it doesn’t give any significant meaning. When collected and divided by categories it shows a better picture of the study object. The qualitative and quantitative analysis helps collecting the right information from unstructured data. By using mining tools the data is processed and divided by importance and the impact it brings to the society or the company.

Sentiment analysis

The opinion mining is that field of study that analyses people’s opinion based on their sentiments and situational behaviors. The emotions and attitudes are part of their life behavior which shows their preferences and their trends. By using this kind of information the output is very helpful for organizations and businesses in retrieving valuable information from client prospect. There are different names for the sentiment analysis such as emotion analysis, sentiment mining and opinion mining, but they all are based on sentiment analysis of opinion mining.

Since 2000 different researches are conducted, based in natural language used to express opinions and sentiments. There are many reasons for the growing of this area. The opinions are applied in most of the domains, the commercial profits are increasing due to clients sentiments and opinions expressed in their feedbacks using social tools. Many problems that have been kept unsolved or studied for years can be treated in a different way.

Due to the technological boom and the growth of social networks, blogs and other social media, now we have collected significant amount of social data with respect to opinions and sentiments. The data is significant in conducting research in social media where using opinions for products and services, organization and individual. These opinions used in the research produce precious information for the object of the study. They show if there is a trend or how strong is a product supported by customers, how reliable is the product etc.

C:\Users\kledia\Downloads\sentiment.jpg

The sentiment analysis is divided in three categories or levels of study. They are based on the type of the data collected and analyzed.

Document level where they classify if the opinion documented is positive or negative. For a given product the sentiment analysis categorizes it based on the public opinions as positive or negative regarding the product. The expressed opinions are only for one product or entity. The system treat each opinion for a single entity and gives the results as positive or negative feedbacks. This category is not applicable for multiple entities in a single document.

Sentence level where the analysis goes to every sentence and evaluates if the sentence expresses a positive or a negative opinion regarding the object of study. When there is a neutral opinion it is evaluates as no opinion at all. The sentence level analysis is related to subjectivity classification (Wiebe, Bruce and O'Hara, 1999), where the subjective sentences are divided from the objective sentences. A difficult part of this level is the misinterpreting of sentiments due to subjective sentences. This type of sentences cannot be interpreted for valuable opinions as they don’t show the real opinion based of facts. Analysts face difficulties in localizing the opinions when they are given embraced with clauses.

Entity and aspect level known also as feature level conducts a better analysis of opinions then the document level and sentence level. The primary objective of this analysis is to find sentiments ad their aspects (overall positive, overall negative, partially positive and negative).this level don’t look in sentences or document to find and locate an opinion as positive, negative or neutral, clauses etc. the opinions is based on sentiments and targets. It directly goes to the opinion classifying it and finding the effects of the opinion. Conducting this type of analysis can be difficult with challenging problems which need to be divided into easy to solve small problems.

Starting from these levels we can create a sentiment analysis for a product where the unstructured text will turn into structured data which will lead to a structured summary of opinions and sentiments. These data can be used for different purpose of data analysis treated as quantitative or qualitative data.

Opinion and Sentiment polls

Public opinion polls are a way to assess / study public opinion, through interviews with a limited sample of individuals, which is considered representative of the population. Methods ranging from direct interviews, those with the phone, or via electronic media, etc.

Public opinion polls are very valuable in public decision-making processes, as used for rapid information collection and study of attitudes / perceptions of society on issues of particular social or political phenomena.

Present perceptions and preferences of the general population.

Test whether a plan, program or element that is presented to the public is acceptable to the public and provide an estimate of the level of preferences or acceptance. For example preference of a candidate or party in the constituency, during an election campaign to predict the chances of their choice.

Test if the opinions are changing, and the dynamics of the reasons for the change. The results can be used as a guide for policy to address the concerns or problems of society, to develop a media strategy, etc.

Focus on public opinion for a service and explain the context for the development or explanation of a certain opinion dominant

Show preferences for certain segments of the population, for example, perceptions of an ethnic or social minority against a law or bill that is being drafted.

Often sentiment analysis is used in politics during elections. People express their opinions in social media, newspapers, magazines, surveys. Their opinions are collected as text where the opinion is given using the natural processing language. After collecting the data will be processed in different tools that sentiment analysis and opinion analysis use. Depending on the tools used the data will be treated as a simple positive, negative or neutral opinion if there is any key word in the text. If the opinion it is not very clear then different system will look at similar words used previously to make any assumption or correlation depending on the ways the opinions are expressed. This is the picture of the process of sentiment analysis.

Individual and group opinions

Usually the object of the study is the human being. Due to his perception for the environment, his ideas and opinions are very important for the growth of the country, the economy and the industry. The aim of this kind of study is to measure and analyze the trends and dynamics of the perception of the population or the target group responses to the phenomena of certain economic, political, social, and consumer behavior towards products / services. Having the results of the study it will be easier to retrieve significant conclusions regarding people preferences of products and services. Getting to know what customers think and expect from a product will lead the company in mitting the client expectations and giving more value to the customers. By focusing in these procedures a new product launched in the market tend to be a success and welcomed by customers.

3.1 Applications & Advantages:

Tracking the dynamics of perceptions or assessments (impact panel)

Advantageous due to the possibility for immediate study of current affairs and attitude towards certain phenomena

Low Cost

Provides the ability to study and analyze the behavior of different population groups (age, different areas of the country, bringing groups with different educational levels, etc.).

Analysis of the perception of the brand (Brand Awareness)

Socio demographic analysis of target groups

Segmentation of target groups

Control the effectiveness of advertising campaigns

Measurement of trends and attitudes of the consumer / citizen on various issues

Research on the image (companies, products, markets)

Study of brand loyalty, frequency of use, purchase, etc.

Research and preferences for media attendance

Analysis of trends, etc.

Opinions

Opinions are categorized as regular and comparative. A regular opinion is knows also an opinion which expresses a point of view, a perception. It can be divided in two parts:

Direct opinion

Indirect opinion

The direct opinion is an opinion that is given in an instant for different reasons. It expresses a direct opinion on a product or entity. The indirect opinion for a product or an entity is expresses indirectly using its effects on other products or entities. Most of the focus is on direct opinions expressed for situations, products and services because they are very simple to analyze. Indirect opinions are more complex and require more time dedicated to analysis in order to give a final result with will be used in different domains.

Comparative opinions are those opinions produced by comparing features of similar products and services. It also looks at different features of similar products, dividing same of similar applications from different applications for the same product, for example comparing applications of Iphone 4s an Iphone 5. By using this analysis we can have a better knowledge what to buy and where to spend money for based on customer’s experience.

Explicit and implicit opinions are another category of opinions used in sentiment analysis. Explicit opinions are subjective phrases that express a regular opinion or a comparative opinion. Explicit opinions are objective phrases that express a desirable or an undesirable product or service. Explicit opinions are easier to find then implicit ones.

The way an opinion is interpreted, depends from the person who is reading the opinion. From the author point of view the opinion can be expressed differently and from the reader point of view can be interpreted with a different meaning. An opinion has different meaning for different category of people. When a price for a product goes down that is a negative impact for the company but a positive impact from a buyer prospective because he can buy more with the same amount of money without taking into consideration what negative impact has occur to the company. Opinions about products can be classified as advertisement for products. If the opinions are positive it will increase customer demand. In the same way if the opinion about a certain product is negative it will lead to a decrease of a customer demand.

Techniques and Tools

5.1 Using supervised learning

The sentiments are divided in positive and negative. The process of analyzing the sentiments is based on texts posted online. Every person can rate different products by using the star application using a scale of one to five stars or one to seven stars. The sentiment analysis is based on the cumulative points/stars that a product collects. Higher the number of stars, higher the sales of that product and customers preferences. Using this way of rating, the analysis becomes easy to conduct and produce valuable results for the industry as well as for customers. To decide if a sentiment is positive or negative we look at the rating of a product, where if the rating is usually 4 or 5 means that that the sentiment expressed is positive otherwise when the rating is below 3 the sentiment is classified as negative. In the case of a product receiving a rating of three stars, the sentiment is classified as neutral and doesn’t show a positive or negative opinion for the customer’s point of view. When trying to retrieve positive or negative opinions from a simple text, the sentiment analysis bases the study into key words like very bad, poor, good, very good and excellent, where these words are being ranked from 1 to 5.

Supervised learning is a technique that can be used in text mining using simple vectors, holding comparative words to classify sentiments in negative and positive groups. Two basic methods are used for sentiment analysis, the Bayes Naïve Classification and Support Vector Machine. There are many ways in differentiating opinions and classifying them using supervised learning method:

Words and their frequency of usage

Sentimental words and statements

Parts of a text

Rules in expressing an opinion

Shifted sentiments

Dependent words

5.2 Using unsupervised learning

The method of unsupervised learning is a technique to classify sentiments expressed by words and statements. These words that are present in different texts are called fixed words. Based on them the algorithm of unsupervised learning starts screening the text and finding possible sentiments expressed in statement of any product or services offered by a company, government etc. The algorithm is goes through three steps:

First step

This phase consists in selecting from a phrase the noun or adjective that describes the adverb of a sentence. So basically the first and the second word of a phrase are eliminated and the third word is selected which can be an adjective that illustrates the condition of the adverb of the sentence. A problem with this type of techniques is the meaning of a word in different contents. A word may express a positive sentiment or opinion in a statement but the same word can express a negative sentiment or opinion in a different statement.

Second step

This phase consists in PMI (pointwise mutual information) where the formula used is:

PMI =

PMI finds if there is any dependency between words. The refers to the number of occurrence of those these terms together in a phrase. The occurrence is given by the probability only if the terms are independent. To measure if a sentiment is positive or negative we use the sentiment orientation formula where:

SO(statement)= PMI( statement, excellent) – PMI (statement, poor)

After retrieving the number of occurrence of a word in a phrase we can try to estimate it alone or associated with another term/word.

Third step

In the first two steps we select and find the probabilities of occurrence of the opinions expressed in a phrase. In the third step we have collected all the data required to estimate the sentiment analysis. We review the pairs of words and compare them to give the type of sentiment and its range. In the final step the algorithm compares all the data and produces an average of occurrences of the sentiments and compiles a generalization for different mediums and domains of usage. The implementation of this method is very useful in ranking movies, cars, technology components and other similar products.

The lexicon method is another unsupervised method where the algorithm uses a dictionary of words and phrases of sentiments with their orientations of strong expressions. This method was previously used in sentences and level sentiments.

The classification of sentiments depends on the domain that they have been used in. In different domains sentiments can have different meaning and express different opinions. The same word can have positive and negative meaning in the same time in two different domains. Many methods have been proposed and tester for sentiments in different domains like the graph approach introduced by Wu, Tan and Chen in 2009 where the documents in the graph are treated as nodes and the link between two notes has a weight using the cosine similarity. Another method created by He, Lin and Alani in 2011 is using joints topic where the sentiments are modeled to identify the topics of sentiments in order to build a bridge between different domains.

Another method included supervised learning by using different beside English to make a comparison of ratings of the same product/service of the same company in different countries. This method was constructed by Boyd-Graber and Restinc in 2010 extending the sentiment analysis beyond borders. They used the topic modeling based on different dictionary languages. One of the last researches conducted by Duh, Fujino and Nagata shows that there is still the problem of matching words in different languages and it is not caused by the machine translation using different dictionaries. The machine translating can produce the proper meaning of the word but the algorithms used to process the data will produce a mismatch between different domains.

With the faster growing of technology and software applications, there will be different approaches in the sentiment analysis and opinion mining field, which will lead the manufacturing and service companies further in another level giving more value to the customers and meeting their expectations.