2.1. NLP Applications

The preprocessing and translation methods to be applied to a corpus depend upon the application being built, and the format of data desired for the application input. In order to identify the preprocessing and translations, the seven classes of NLP applications and their goals have to be well understood:

• Sentiment Analysis. In SA, an input corpus is analyzed to determine a sentiment on some finite continuous scale. For example, the corpus may be analyzed and assessed as to whether or not the text conveys a positive message (on the +ve scale) or negative message (on the -ve scale.) The end goal is to take any previously unseen input and determine an accurate measure of sentiment for that input. This is a generally a supervised application.

• Text Prediction (TP). In TP, the input corpus is interpreted and then an attempt is made to produce the terms that would follow an incomplete input message. For example, if the input were “I like to watch the rain ____,” where the underscore represents the incomplete end to the message, the model should predict “fall” or “drip,” or another appropriate word, to complete the message.

• Analogy Completion (AC). In AC, for example, the input of three words representing three of four terms in a pair of analogies is taken. If the input was “kitten” is to “cat” as “puppy” is to ____, where the underscore represents the prediction the model will make, the model should predict “dog” to complete the analogy.

• Text Classification (TC). In TC, an input corpus such as a document or other stream of text is used to attempt to identify the input as belonging to one or more classes of information. For example, if the goal of the model is to classify job postings into one of thirty industries, it might take an input describing a cybersecurity technician position as belonging to the “Information Technology” industry.

• Document Summarization (DS). In DS, a large input corpus is used to interpret the meaning, and produce a much smaller output corpus which attempts to accurately summarize the original corpus. For example, many DS applications are designed to read research papers and produce an accurate abstract or summary for readers to skim or use to understand the relevance and usefulness of a paper to their research.

• Machine Translation (MT). In MT, a stream of input corpus is used to interpret the meaning, style, and form of speech, and attempt to produce a translation to another language or representation. For example, many MT applications are designed to translate an input stream from one language to another, like English to Russian.

• Question Answering (QA). In QA, a short input corpus is taken for asking a directed question, analyzing the meaning and intended response, and attempting to produce a suitable answer to the question posed. For example, if asked “Where can you buy groceries?,” the model should reply with “The grocery store” or something similarly suitable.

2.2. Dataset Translations and Transforms

Multiple translation/transformation methods may generally be used for each of the NLP application types; for an estimate of these, see Table 1. Popular translation and transformation methods are explained below.

• Word embedding. For WE, an input corpus of tokens is translated by translating each token into a multi-real-valued vector encoding the meaning of the word relative to a selection of words in the English language. In WE, words that convey similar meaning are located relatively close to one another in high-dimensional space; words with the same meaning but opposite connotation may exist on a high-dimensional line representing the scale of sentiment.

The dimensions of meaning in WE are abstract, but methods such as Term Frequency/Inverse Document Frequency (TF-IDF) combined with t-Distributed Stochastic Neighbor Embedding (t-SNE) produce a WE library of a language and then represent it in lower-dimensional space for human-comprehensible coarse interpretation, respectively.

WE representation can be used to compare the meaning of words, and later to translate an input corpus to the average of the embeddings of its terms, producing an approximate encoding of meaning.

• Binary classification (BC). In BC, the problem presented requires an output that segments an input into two distinct and typically orthogonal classes. BC is perhaps the most simple transformation, as each term or group of terms is classified into one class or another (generally represented ss 0 or 1, with an output on the continuous scale 0~1.)

• Multi-class classification (MC). When a problem requires an approximate classification of the input into a distribution of probabilities for its class, it falls under the category of MC. In MC, an input, whether supervised, semi-supervised, reinforcement, or other actor-critic training methods are used, is denoted as belonging with certain probabilities to each of the proposed output classes.

For example, taking the classification of job listings from a previous example, the job listing for a cybersecurity technician position may fall under multiple industry categories: “Computer Science,” “Information Technology,” “Information Systems,” “Tech Support.” Each category has its own likelihood of accuracy, and so may be represented in supervised learning or proposed in the output of the model. This job listing may be 0.9 likely to be IT, 0.08 likely to be CS, and 0.02 likely to be IS, but not likely to be Tech Support.

• One-hot encoding (OH). OH is typically used similarly to MC, but when the desired output is a single class with more orthogonal meaning or representation. OH represents a classification as a 0 in all places of an output vector, with a single 1 in a single space encoding the desired classification.

In the former example, the application may need to strongly delineate job listings into a single industry in order to properly index and return listings upon user searches. In this case, OH would be applied, for example, in supervised learning, to tag an input with a single class identifying its expected industry.

• Part-of-speech tagging (POST). POST is generally used in NLP applications that intend to perform deeper analysis of meaning of an input corpus, such as in QA and TP. POST transforms each word in the input corpus to its respective part of speech: an adjective, verb, noun, pronoun, stop word/notation, and so on.

POST allows for the preservation of tense, and is typically combined with WE methods when used for QA or TP in order to produce an output consistent with the tense and writing style of the input.

Table 1 presents a summary of the goals, translations/transformations, and methods.

2.3. WEs as a Focus

This work focuses on WEs as a method for producing rich sentimental data out of the limited corpora contained in each of the selected tweets. WE was selected since WE captures a broader and deeper semantic meaning in the full scope of language. This representation of meaning is used to determine a singular value for the sentiment that is based on axes of varying meanings rather than a simple barometer of “positive” and “negative.”

3.1. Works on Preprocessing

Several works have also been done on preprocessing online texts [4–7]. Saif et al. [5] looked at the effectiveness of removing stop words in the context of Twitter sentiment classification. In this paper they investigated whether removing stop words helps or hampers the effectiveness of Twitter sentiment classification methods. They recommended that the dynamic generation of stopword lists, by removing infrequent items appearing only once in the corpus, appears to be the optimal method to maintaining a high classification performance. Haddi et al. [4] explored the impact of preprocessing methods on Twitter sentiment classification. Bao et al. [7] looked at the effects of URLs, negation, repeated letters, stemming, and lemmatization. Their results show that classification accuracy rises when negation transformations are performed and repeated letter normalization is employed, and descends when stemming and lemmatization are applied.

3.2. Works on Quantifying Emotional Polarization

There are many papers on SA exploring both traditional and machine learning approaches to NLP as they are applicable to quantifying emotional polarization. In tandem with the popularity of social networks, SA has increasingly become an area of interest for many studies.

Akaichi et al. [8] collected some comments from Facebook users during the Tunisian Revolution. Emotional analysis of corpus was performed using machine learning algorithms. The Naive Bayes (NB) and support vector machine (SVM) algorithms were compared by combining different feature extractors. Higher accuracy was obtained by combining different characteristics for SA. They constructed a sentiment lexicon, based on emoticons, interjects, and acronyms.

Gonçalves et al. [9]'s study aims to compare the popular methods to find out the potential limitations, advantages, and disadvantages for SA. The researchers compared eight popular SA methods in terms of accuracy and agreement to classify the sentiment of each input text. Also, they developed a new method that combined existing approaches, providing the highest accuracy and comparative agreement. They use positive or negative to identify polarity.

Vishal and Sonawane [10] gathered and labeled a collection of positive, negative, and neutral tweets. A number of linguistic statistical analyses were performed on the collected corpus. The NB classifier and SVM were used to classify each text. The results show that the NB classifier has a better classification effect. When training the Bayesian classifier, two different approaches were implemented: one of them combined the presence of n-grams and the other combined the part-of-speech distribution information. Their results indicated that the best performance was achieved using bigrams.

Hamid and Islam [11] analyze the collection of tweets in two ways. First by a method of literary analysis to classify the emotion of each tweet using a lexicon-based approach, and second, using a machine learning-based approach. Machine learning methods in NLP include Case-Based Reasoning, Artificial Neural Networks, and SVMs. They measured tweets in terms of positive, negative, and neutral.

Erik and White [12] elaborate on the development of NLP technology, including the historical background of NLP technology and some likely evolutions. A semantic-based NLP method was introduced.

Rajman and Besançon [13] studied the methods, applications, and techniques of text mining. It was demonstrated that data mining technology can extract useful information from text-based databases. According to different sources of information extraction, the authors propose four methods of text mining, which are term-based methods, paragraph-based methods, concept-based methods, and pattern classification methods. Useful results of text mining are explored, including the generation of association rules.

Agarwal et al. [14] uses linguistic features to perform SA on Twitter data. The experiments on Twitter SA showed that part-of-speech features may not be useful for SA in the microblogging domain. More research is needed to determine whether the POS features are just of poor quality due to the results of the tagger or whether POS features are just less useful for SA in this domain. Features from an existing sentiment lexicon were somewhat useful in conjunction with microblogging features, but the microblogging features (i.e., the presence of intensifiers and positive/negative/neutral emoticons and abbreviations) were clearly the most useful.

Luo et al. [15] also looked at a negative or positive opinion about a particular topic. They used social features to improve performance.

3.3. Works on Political Content Analysis on Twitter

Few works have looked at the political analysis of Twitter data. Jungherr [16] provided a comprehensive literature review of 115 studies on Twitter in Politics. Small [17] looked at the Canadian political aspects of Twitter data in relation to the hashtag. Ausserhofer and Maireder [18] looked at the interrelation of individuals on the basis of their professions and their connection to mass media using Austrian political data.

Most of the works presented in this section measure polarity by a single axis of positivity and negativity. Our work attempts to study microblogging corpuses using multiple axes, to better express the sentiment of the political tweets.

5.1. Preprocessing

SAs of Twitter corpuses will be highly affected by the inherently noisy nature of Twitter data, hence to analyze a corpus of tweets, the first step is preprocessing. Preprocessing involves a series of techniques, presented below, mainly performed to get better results. First general preprocessing methods for Twitter corpuses are presented, and then the preprocessing done on the Twitter dataset used in this paper is presented.

5.1.2. Preprocessing the Twitter data

Before performing word segmentation, initial preprocessing was performed on the CSV file, for example, removing punctuations, numbers and stop words. English words in uppercase were converted into lowercase. Abbreviations and contractions had to be restored. For example, some common contractions had to be resolved: “aren't”: “are not,” “can't”: “cannot,” “couldn't”: “could not,” “didn't”: “did not,” “doesn't”: “does not,” “don't:” “do not.” Expanding these contractions helped in the subsequent segmentation of words.

Stop words are words that appear often but have little meaning. In English, stop words are, for example, “a,” “and,” “is,” “the,” “of,” and so on. For this analysis, text analysis was being performed based on word frequency and statistical methods, and stop words would reduce the accuracy of the analysis, hence were removed.

There were also many punctuation marks, user mentions, multimedia, and links in the original data. These could not be handled by the proposed model, and hence were removed. At the conclusion of preprocessing, only English words that were of sentimental value were left. Future work will return to applying the usefulness of rich non-English sentiment indicators as mentioned in [14] to produce more robust representations of microblogging text such as tweets for analysis as performed in this paper; for simplicity of proposing a novel approach to SA, these tokens were dropped.

Finally, word segmentation was performed. Word segmentation splits the input text into each token separated by a space. The result is the final input. This process is illustrated in Figure 1.

Figure 1

Illustration of preprocessing.

5.2. Calculating Polarization

The proposed method builds off of prior work in WEs, social networks, SA, and linguistics, to efficiently calculate a per-word and per-corpus sentiment value easily fine-tuned to the problem space. To explain the proposed method, first the format is explained, and then the implementation, and value of WEs is presented. WEs are high-dimensional representations of all of the recognized words in a language (sometimes including misspellings, generalized phrases with numbers or percentages, etc.) with values determined as a product of the use of the word in different contexts.

For example, the word “burger” will likely appear frequently in statements about food, that is, “I want a burger.” The “closeness” of the word to the other words in context is used to slowly approximate the “meaning” of the word in high-dimensional space; words further away should be further away from the target word, as they carry different meaning. Ultimately, other “food words” like “smoothie” will tend to appear in similar contexts, and gradually move in high-dimensional space to be relatively close to “burger.” The groups that form are easily recognized by human interpreters, but WE is the most effective method that machine learning models have been able to use to thoroughly and efficiently interpret natural language.

SA, as target of this proposed method, has important applications in the interpretation of natural language [19], particularly on social media. SA broadly means the classification of a corpus into a positive or negative category based on the meaning of the words it contains. For example, “I hate shiny fountain pens” should, with a well-trained model, classify as a negative statement. However, in this particular case, it is also important to note the small corpus, and the minor sentimental value in the other words: the pronoun is unlikely to carry any value, while “fountain pens” may contain a strong positive value as they are typically referenced in positive contexts, for example, “I love shiny fountain pens.” The corpus may, for that reason, be rated neutral or slightly positive, when in fact the primary acting word “hate” should drive the sentiment to negative. The model has associated the sentiment of “fountain pens” out of context, and thus spoiled the overall sentimental meaning.

This problem is exacerbated on microblogging social media platforms like Twitter, where the corpus to classify is typically a small number of words combined with hashtags, emoticons, and emojis references to other users, and typically incorporate multimedia like photos or videos. A short corpus, or one that does not contain as many words as it does other indicators of meaning, are hard for NLP models to classify. The above example is made worse with these datasets where a limited number of words in the corpus carry sentimental value and can accidentally steer the sentiment toward the wrong axis.

In order to combine these two conclusions and create a model that is resistant to misclassification of sentiment and factors in the high-dimensional meaning of words, this proposed generalized model for SA of social media corpora is presented.

5.2.2. Cosine similarity

CS is a standard technique used in WE for comparing the similarity of two words, that is, embeddings [20,21]. CS is equivalently the dot product over the product of the magnitudes:

similarity=cosθ=A⋅B|A||B|

This produces a calculation of the distance between any-dimensional vectors. In most Word2Vec models, including the model used in this paper, this would be vectors with 300 dimensions. Each vector is normalized before calculating the CS, and the result constitutes a measure of “similarity” of the two vectors from −1.0 to 1.0 in how closely the two vectors represent the same direction in their dimensional space, as shown in Figure 2.

Figure 2

Calculating cosine similarity.

In addition to using CS to determine the polarization value of each individual token, the mean value of each dimension of the embeddings is also taken to be the encoded meaning of the entire selection of text. The absolute value CS of this vector and the polarization axis are then calculated. This value presents the overall distance of the selection to the sentiment axis—agnostic of its True Positive (TP) or True Negative (TN) value—which demonstrates how closely the selection is polarized along the axis. A selection of text which is heavily polarized in either direction indicates that the axis is accurate for the corpus; a selection of text which does not clearly align with the axis may indicate that the axis is not an accurate polarization of the corpus.

The average of this value was taken for each text in the corpus to be a representation of how accurately the selected axis represents the polarization of the corpus, deemed here the axis score (AS). How this value helps in determining the appropriate axis for any given corpus is discussed further in the experimental results. The results may contribute to future work in accurately analyzing the polarization for a subject discussed on social media platforms in general and microblogging platforms in particular. An axis with a higher AS implies that the axis more accurately represents the polarization of the corpus with respect to the subject. The AS is calculated as follows:

Axis Score=1n∑i=1ncsaxis,veci

where n is the number of texts, axis is the selected axis of polarization, and veci is the embedding of mean values for the i-th text.

Figure 3 demonstrates a chosen axis and a selection of WEs which result in a low overall AS. The absolute value of the CS of V1 is low, indicating its meaning is not closely related to the sentiment captured by the axis and is therefore of little value in determining sentimental value with respect to the axis; the CS of V2 is relatively high, indicating that its meaning is closely related to the sentiment captured by the axis; and V3 is orthogonal to the axis with a CS of 0, indicating it captures an exactly opposite meaning than the axis (which while improbable, serves the purpose of example).

Figure 3

A set of disjoint word or corpus (V1-3) embeddings compared to the axis of sentiment (A) and their associated low cosine similarity (CS) values CS(A, V#).

Figure 4 demonstrates a chosen axis and a selection of WEs which result in a high overall AS. The absolute value of the CS of all three vectors is high but not exact, indicating all three words meaningfully contribute to determining sentimental value with respect to the axis. If a vector were to have an absolute CS of 1, indicating an exact match with the axis, it is almost certain that the word is in fact the same as the word chosen to determine the axis in that direction.

Figure 4

A set of cohesive word or corpus embeddings with high cosine similarity (CS) values.

The mean polarization value of the tokens of each text is taken and compared with the BC of this value against a hand-labeled validation set. An axis with a more evenly distributed truth table—in the case of microblogging platforms, a result which should yield a relatively even likelihood of true or False Positive (FP) or False Negative (FN) classifications—is more likely to be an accurate polarization of the corpus with respect to the subject. A truth table skewed toward positive or negative classification indicates a poorly selected axis where the consensus of polarization on the subject tends toward one side, and thus the true polarization should be more specific or better tuned.

These two approaches are combined to analyze the polarization axis to select the most appropriate axis for the subject.

5.3. Methodology

The proposed method, as applied to this dataset, depicted in Figure 5, is presented next.

1. Identify the topic of analysis and a word representing the positive and negative sentiment for it.

   Example: Impeachment; “guilty” versus “innocent”

2. Take the high-dimensional vector from the WE of the negative word to the positive word as the axis of polarization using a trained Word2Vec model.

   Example: model.embedding[“innocent”] - model.embedding[“guilty”]

3. Collect corpora to classify and analyze.

   Example: Unique tweets containing the word “impeach” from the Twitter API

4. Perform preprocessing on each corpus as described in the Preprocessing the Data section.

5. For each input:

   1. Split the input into its individual tokens.

   2. Use the Word2Vec model to translate each token into its WE.

      Calculate the CS of each token to the polarization axis to determine its positive or negative polarization value.

   3. Take the mean of the polarization of all tokens in the input as the polarization value of the whole input.

   4. Use a rounding scheme to determine a binary positive or negative classification for the input ascribed to our fine-tuned positive and negative meaning.

Figure 5

Experimental design.

6.1. Single Axis

Table 2 presents the results of each single axis, A1 to A13. Affirmative values are bolded. For this work, affirmative values were considered to be those of an accuracy close to 50%, CB less than 0.25, and AS above 0.25. The remaining values are taken to indicate rejection of the axis.

From these results it is evident that there is not a clear optimal axis among the proposed axes A1 through A13. Though most axes maintained an accuracy around 50%, this is not a high classification accuracy. While A3 (negative-positive) has a relatively neutral CB, it has a low AS; A4 and A5 have high AS, but extreme negative CB values.

While AS can indicate a finely tuned polarization axis, CB will also be prioritized in identifying the most appropriate axis. Thus, in this case, the standard positive-negative sentiment axis A3 remains the most appropriate axis for analysis.

It is significant that for each of the proposed axes except for A3, the chosen axis words are present in the corpus with significant probability. “Positive” and “negative,” however, do not appear frequently, and thus there is less influence on the calculation of polarization for each input in the absence of either word.

6.2. Combined Axes

In order to improve the metrics of the results of the single or individual axes, the averages of multiple axes were used. This is the vector represented by the mean of values in each dimension of the axes selected. From the single axes in Table 2, the combined-axis classifications were performed, as presented in Table 3, which are reasonably related but not synonymous, and also showed meaningful or promising potential individually.

One combination of axes, A3 + A6, improves significantly on the AS of each individual axis, while also approaching the mean of the accuracy and CB values, weighted heavily toward the A6 axis' original CB. Actually, the improved performance on these metrics for the (A3 + A6) combination could be preemptively identified by noting that A3 showed a slightly positive CB while A6 was slightly negative, and thus their combination tends toward the middle; the AS improves significantly, from which it can be inferred that the sentiment expressed in these tweets can be extrapolated as a combination of feelings of “negative” and “liberal” versus “positive” and “conservative.” These initial results from combining axes indicates that there may be more potential in exploring multi-faceted sentiment from combining more or different axes.