Text analysis is the automated process of obtaining information from text.
In today's information-saturated world, it's a challenge for businesses to keep on top of all the tweets, emails, product feedback and support tickets that pour in every day. Take Google, for example. On average, the tech company processes over 40,000 search queries every second, which is equal to over 3.5 billion searches per day!
So, how can text analysis help businesses deal with information overload?
Below, we'll go into more detail about what text analysis is, how it works, use cases and applications, as well as some resources and useful tutorials to get your feet wet. Maybe you're new to artificial intelligence and work in customer support, sales or product teams. You might even be a data-savvy analyst or software developer. Either way, this guide offers a comprehensive introduction to text analysis with machine learning.
Read this guide in your spare time, bookmark it for later, or jump to the sections that pique your interest:
Let's get down to the nitty-gritty!
Introduction to Text Analysis
What is Text Analysis?
In short, text analysis (a.k.a. text mining) is the automated process that allows machines to extract and classify information from text, such as tweets, emails, support tickets, product reviews, survey responses, etc.
Businesses might want to extract specific information, like keywords, names, or company information. They may even want to categorize text with tags according to topic or viewpoint, or classify it as positive or negative.
Either way, sorting through data is a repetitive, time-consuming and expensive process if done by humans – just imagine if Walmart's employees had to manually process the one-million customer transactions they receive every day. It would take forever. Instead, if done by machines, high volumes of text can automatically be analyzed, saving time and money, providing more insights from business data and automating processes.
To really understand what automated text analysis is, we need to touch upon machine learning. Let's start with this definition from Machine Learning by Tom Mitchell:
"A computer program is said to learn to perform a task T from experience E".
In other words, if we want text analysis software to perform desired tasks, we need to teach machine learning algorithms how to analyze, understand and derive meaning from text. But how? The simple answer is by tagging examples of text. Once a machine has enough examples of tagged text to work with, algorithms are able to start differentiating and making associations between pieces of text, and can even begin to make predictions.
It's very similar to the way humans learn how to differentiate between topics, objects, and emotions. Let's say we have urgent and low priority issues to deal with. We don't instinctively know the difference between them – we learn gradually by associating urgency with certain expressions. For example, when we want to identify urgent issues, we'd look out for expressions like 'please help me ASAP!' or 'urgent: can't enter the platform, the system is DOWN!!'. On the other hand, when we want to identify low priority issues, we'd look out for more positive expressions like 'thanks for the help! Really appreciate it' or 'the new feature works like a dream'.
The number of texts needed for training a machine learning model will depend on various factors such as the complexity of what you want to achieve, and the number of tags involved.
- Topic detection needs around 250 examples per tag (topic) for good accuracy levels.
- Sentiment analysis needs at least 500 examples per tag (sentiment) to produce good results.
These are just two exciting capabilities of text analysis, which we'll come back to later on. First, let's explore some definitions that are closely linked to text analysis.
Text Analysis vs. Text Mining vs. Text Analytics
Firstly, let's dispel the myth that text mining and text analysis are two different processes. The terms are often used interchangeably to explain the same process of obtaining data through statistical pattern learning. To avoid any confusion here, let's stick to text analysis.
So, text analytics vs. text analysis: what's the difference?
One is qualitative and the other quantitative, the latter being text analytics. If a machine performs text analysis, it identifies important information within the text itself, but if it performs text analytics it reveals patterns across thousands of text data, resulting in graphs, reports, tables etc.
Here's an example:
Let's say a customer support manager wants to know the outcomes of each support ticket handled by individual team members – was the result positive or negative? By analyzing the text within each ticket, and subsequent exchanges, customer support managers can see team members' individual ticket resolution rates. However, it's likely that the manager also wants to create a graph that visualizes how many tickets were tagged as solved. They'd use text analytics.
Basically, the challenge in text analysis is decoding the ambiguity of human language, while in text analytics it's detecting patterns and trends from the results. As in the example above, text analytics and text analysis are often used hand in hand.
Methods and Techniques
There are basic methods for text analysis and more advanced ones. First, let's start with the simpler techniques.
Word frequency can be used to list the most frequently occurring words or concepts in a given text. This can be useful for a number of use cases, for example, to analyze the words or expressions customers use most frequently in support conversations, e.g. if the word 'delivery' appears most often, this might suggest there are issues with a companies delivery service.
Collocation helps identify words that commonly co-occur. For example, in customer reviews on a hotel booking website, the words 'air' and 'conditioning' are more likely to co-occur rather than appear individually. Bigrams (two adjacent words e.g. 'air conditioning' or 'customer support') and trigrams (three adjacent words e.g. 'out of office' or 'to be continued') are the most common types of collocation you'll need to look out for.
Collocation can be helpful to identify hidden semantic structures and improve the granularity of the insights by counting bigrams and trigrams as one word.
Concordance helps identify the context and instances of words or a set of words. For example, the following is the concordance of the word simple in a set of app reviews:
In this case, the concordance of the word simple can give us a quick grasp of how reviewers are using this word. It can also be used to decode the ambiguity of the human language to a certain extent, by looking at how words are used in different contexts, as well as being able to analyze more complex phrases.
Now that we've touched upon the basic techniques of text analysis, we'll introduce you to the more advanced methods: text classification and text extraction.
Text classification is the process of assigning predefined tags or categories to unstructured text. It's considered one of the most useful Natural Language Processing (NLP) techniques because it's so versatile and can organize, structure and categorize pretty much anything to deliver meaningful data and solve problems.
Below, we're going to focus on some of the most common text classification tasks, which include sentiment analysis, topic modeling, language detection, and intent detection.
Emotions are essential to effective communication between humans, so if we want machines to handle texts in the same way, we need teach them how to detect emotions and classify text as positive, negative or neutral. That's where sentiment analysis comes into play. It's the automated process of understanding an opinion about a given subject from written or spoken language.
For example, by using sentiment analysis companies are able to flag complaints or urgent requests, so they can be dealt with immediately – and perhaps avert a PR crisis on social media. Other uses of sentiment classifiers include assessing brand reputation, carrying out market research, and improving products with customer feedback.
Why not play around with our pre-trained classifier using MonkeyLearn? Write something positive or negative, and see how this classifier makes a prediction:
For more accuracy, you can make your own custom classifier for your specific use case and criteria. Check out these use cases & applications to see how companies and organizations are already using sentiment analysis.
Another common example of text classification is topic analysis or, more simply put, understanding what a given text is talking about. It's often used for structuring and organizing data. For example:
“The app is really simple and easy to use”
This product feedback can be classified under Ease of use.
Language detection allows text to be classified according to its language and is often used for routing purposes e.g. support tickets can be automatically allocated to the right team depending on the language detected. Saving time and avoiding confusion among staff.
We've created a classifier that detects 49 different languages in a text. Discover the language of the text below:
Text classifiers can also be used to automatically detect the intent within texts, for example, companies are able to better understand customer feedback about a product if they know more about the purpose or intentions behind the text. Anything from the intention to complain about a product to the intention to buy a product.
Check out the following classifier, trained for detecting the intent from replies in outbound sales emails. We've used the tags interested, not interested, unsubscribe, wrong person, email bounce, and -autoresponder to train this classifier:
Text extraction is another widely used text analysis technique for getting insights from data. It involves extracting pieces of data that already exist within any given text, so if you wanted to extract important data such as keywords, prices, company names, and product specifications, you'd train an extraction model to automatically detect this information. Then, you can organize the extracted data into spreadsheets, translate into graphs and use it to resolve particular problems. And yes, all without having to tediously sort through data, and input information manually!
Text extraction is often used alongside text classification so that businesses can categorize their data and extract information at the same time. There are different extraction models for different types of purposes, which we'll go into more detail about below.
Keywords are the most relevant terms within a text, terms that summarize the contents of text in list form. Keyword extraction can be used to index data to be searched and to generate tag clouds (a visual representation of text data).
Use the following keyword extractor to see how it works:
Remember, if you'd like more accurate results you'e better off training your own custom extractor and tailoring it to your particular needs and criteria.
A named entity recognition (NER) extractor finds entities, which can be people, companies or locations and exist within text data. Results are shown labeled with the corresponding entity label, like in this pre-trained person extractor:
Summary extraction allows long texts to be summarized without losing their meaning. Imagine if every customer support ticket could be read in seconds instead of minutes – staff would save a lot of time, and response times would be quicker leading to better customer experiences.
See for yourself how this summary extractor handles a piece of text:
Above, we've mentioned the most common models of text analysis, but there are many other useful methods. Below, we've listed two more advanced techniques that we think deserve an honorable mention!
Word Sense Disambiguation
It's very common for a word to have more than one meaning, which is why Word Sense Disambiguation is a major challenge of Natural Language Processing. Its function is to identify the sense of a word within a sentence when the word has more than one meaning. Easy for humans to figure out, but difficult for machines.
Take the word 'light' for example. Is the text referring to weight, color or an electrical appliance? Smart text analysis with Word Sense Disambiguation can differentiate words that have more than one meaning, but only if we teach models to do so.
Text clusters are able to understand and group vast quantities of unstructured data. Although less accurate than classification algorithms, clustering algorithms are faster to implement because you don't need to tag examples to train models. That means these smart algorithms mine information and make predictions without the use of training data, otherwise known as unsupervised machine learning.
Google is a great example of how clustering works. When you search for a term on Google, have you ever wondered how it takes just seconds to pull up relevant results? Google's algorithm breaks down unstructured data from web pages and groups pages into clusters around a set of similar words or n-grams (all possible combinations of adjacent words or letters in a text). So, the pages from the cluster that contains a higher count of words or n-grams relevant to the search query will appear first within the results.
Text Analysis Scope
Text analysis can stretch it's AI wings across a range of texts depending on the results you desire. It can be applied to:
- Whole documents: obtains information from a complete document or paragraph e.g. the overall sentiment of a customer review.
- Single sentences: obtains information from specific sentences e.g. more detailed sentiments of every sentence of a customer review.
- Sub-sentences: obtains information from sub-expressions within a sentence e.g. the underlying sentiments of every opinion unit of a customer review.
Why is Text Analysis Important?
Every minute of the day, 156 million emails and 456,000 tweets are sent! That's a colossal amount of data to process, and impossible for humans to do it alone. If machines are made solely responsible for sorting through data using text analysis models, the benefits for businesses will be huge.
Let's take a look at some of the advantages for businesses below:
Text Analysis allows businesses to structure vast quantities of information, like emails, chats, social media, support tickets, documents and so on, in seconds rather than in days, and redirect extra resources to more important business tasks.
Businesses are inundated with information, making it harder to resolve urgent queries and deal with negative reviews as and when they arise! Text analysis is a game-changer when it comes to detecting urgent matters, and the big advantage is that it can work in real-time 24/7. By training these models to detect expressions and sentiments that imply negativity or urgency, businesses can automatically flag tweets, reviews, videos, tickets, and the like, and take action sooner rather than later.
Humans make errors. Fact. And the more tedious and time-consuming a task is, the more errors that are made. By using automated text analysis models that have been trained, algorithms are able to analyze, understand, and sort through data more accurately than humans. We are influenced by personal experiences, thoughts,and beliefs when reading texts, whereas algorithms are influenced by the information they've received. By applying the same criteria to analyze all data, algorithms are able to deliver more consistent and reliable data.
Throughout this section, we've covered what is text analysis, different techniques, and scopes. We went through the benefits of automated text analysis and learned about text classification and text extraction – that companies can use and customize depending on their business needs.
Organizations that adopt text analysis will have clear advantages over those that don't, including automating business processes, getting actionable insights for better decision making, and processing data at scale and in a cost-effective way.
Now that you know the ins and outs of text analysis, check out our next chapter How Text Analysis Works to learn more about training text analysis models with machine learning. Before you know it, you'll be creating your own text classifiers and extractors!
How does Text Analysis work?
Machine learning for text analysis makes it possible to process huge amounts of unstructured text data in a fast and simple way.
By creating customized models that learn from examples and improve over time, businesses can automate daily tasks and save their teams precious time, as well as gain relevant insights that enhance the decision-making process.
But how easy is to get started? Automated text analysis may sound far too complex for someone with no programming skills, but that's not always true. By using an AI platform like MonkeyLearn, everyone is able to create customized text analysis models or even use pre-trained models for specific purposes, without writing a single line of code!
However, so that businesses can take full advantage of automated text analysis, it's important to understand how it works.
In this section, we'll dive deeper into how text analysis works. First, we'll go through the internal and external sources you can use to obtain data. We'll also refer to the different natural language processing techniques ― like parsing, tokenization, lemmatization, and stop-word removal ― to prepare the data for text analysis.
Finally, we'll provide some insights on data visualization and highlight the different tools you can use to show the results of text analysis.
Before getting started with automated text analysis, we need to gather text data. That could be any data we're interested in analyzing.
For example, let's imagine that we work for Slack and we want to analyze online reviews to better understand what our customers like and dislike about our platform. First, we'd need to automatically gather reviews from sites like Capterra and G2Crowd.
Gathering data is important because it's used as training samples to build either text classification or text extraction models with machine learning. After training models, we'll be able to automatically analyze text and make predictions on our data.
Sources for data gathering can either be internal or external:
It's the data that you already have in your company database, or that you can obtain from the tools you use every day. Every business generates a myriad of unstructured data, from emails and chats, to reviews, customer queries, and support tickets.
In order to use this internal data, you will need to export it from your software or platform as a CSV or Excel file, or retrieve the data with an API.
Here are some examples of where you might get internal data:
Customer Service Software: the software you use to communicate with customers, manage user queries coming from different channels and deal with customer support issues. Zendesk, Freshdesk, and Help Scout are a few examples of this type of software.
CRM: it's the software that allows your company to keep track of all the interactions with clients or potential clients. It can involve different areas, from customer support to sales and marketing. Hubspot, Salesforce, and Pipedrive are some examples of CRMs.
Chat: the apps that you use to communicate with the members of your team or your customers. Slack, Hipchat, Intercom, and Drift are four good examples.
Email: the king of business communication, emails are still the most popular tool to manage conversations with customers and team members. Gmail is the obvious reference when it comes to email service providers, but there are other email management platforms like Front, which is focused on a shared email inbox for teams.
Spreadsheets: they are widely used by companies to compile and analyze data, create reports and budgets, and organize tasks. Whether it's Google Sheets, Excel files or CSV documents, spreadsheets are a valuable source of information.
Surveys: a great way to obtain data, generally used to gather customer service feedback, product feedback or to conduct market research. Some popular survey tools include Typeform, Google Forms, and SurveyMonkey.
NPS (Net Promoter Score): one of the most popular metrics for customer experience in the world. Many companies use NPS tracking software to collect and analyze feedback from their customers. A few examples are Delighted, Promoter.io and Satismeter.
Databases: a database is a collection of information. By using a database management system, a company can store, manage and analyze all sorts of data. Examples of databases include Postgres, MongoDB, and MySQL.
Product Analytics: it's the product feedback and information about the interactions of a customer with your product or service. It's useful to understand the customer's journey and make data-driven decisions. ProductBoard and UserVoice are two tools you can use to process product analytics.
This refers to data that you can find anywhere on the Web. You can use web scraping tools, APIs and open datasets to collect external data from different websites and analyze it with a machine learning model. Let's go into more detail below
Web Scraping Tools:
Visual Web Scraping Tools: these allow you to easily build your own web scraper without needing any coding skills. Dexi.io, Portia and ParseHub are some tools that you can use to crawl a website and extract relevant data without using any code.
Web Scraping Frameworks: if you are a seasoned coder, you can benefit from these tools to create your own scraper and efficiently obtain data from a website. Scrapy, for example, is an open source tool you can use with Python. Wombat is also a powerful scraping tool written in Ruby.
If you are a developer, you can use APIs to connect with different websites and social media platforms and obtain useful data. Facebook, Twitter, and Instagram, for example, have their own APIs and allow you to extract data from their platforms. Major media outlets like the New York Times or The Guardian also have their own APIs and you can use them to search their archive or gather users' comments, among other things.
This is free data from the Web, which can be used for any purpose. You can search for data by topic on websites like Kaggle and Quandl.
In order to automatically analyze text with machine learning, there is some preparation data needs to go through. Most of this is done automatically and you won't even notice it's being done. However, it's important to understand that automatic text analysis makes use of a number of Natural Language Processing techniques (NLP) and processes that you might be unaware of. The output of these processes will help build the input of the machine learning models you create to analyze your data.
Tokenization, Part-of-speech Tagging, and Parsing
The first thing a text analysis system needs to take care of is recognizing what units it will analyze. This is known as tokenization. In other words, tokenization refers to the process of breaking up a string of characters into semantically meaningful parts that can be analyzed (e.g., words) while discarding meaningless chunks (e.g. whitespaces).
The examples below show two different ways in which one could tokenize the string 'Analyzing text is not that hard'.
(Incorrect): Analyzing text is not that hard. = [“Analyz”, “ing text”, “is n”, “ot that”, “hard.”]
(Correct): Analyzing text is not that hard. = [“Analyzing”, “text”, “is”, “not”, “that”, “hard”, “.”]
Once the tokens have been recognized, it's time to categorize them. Part-of-speech tagging refers to the process of assigning a grammatical category, such as noun, verb, etc. to the tokens that have been detected.
Here are the PoS tags of the tokens from the sentence above:
“Analyzing”: VERB, “text”: NOUN, “is”: VERB, “not”: ADV, “that”: ADV, “hard”: ADJ, “.”: PUNCT
With all the categorized tokens and a language model (i.e. a grammar), the system can now create more complex representations of the texts it will analyze. This process is known as parsing. In other words, parsing refers to the process of determining the syntactic structure of a text. To do this, the parsing algorithm makes use of a grammar of the language the text has been written in. Different representations will result from the parsing of the same text with different grammars.
The examples below show the dependency and constituency representations of the sentence 'Analyzing text is not that hard'.
Dependency grammars can be defined as grammars that establish directed relations between the words of sentences. Dependency parsing is the process of using a dependency grammar to determine the syntactic structure of a sentence:
Constituency Phrase Structure Grammars model syntactic structures by making use of abstract nodes associated to words and other abstract categories (depending on the type of grammar) and undirected relations between them. Constituency parsing refers to the process of using a constituency grammar to determine the syntactic structure of a sentence:
As you can see in the images above, the output of the parsing algorithms contains a great deal of information which can help you understand the syntactic (and some of the semantic) complexity of the text you intend to analyze.
Depending on the problem at hand, you might want to try different parsing strategies and techniques. However, at present, dependency parsing seems to outperform other approaches.
Lemmatization and Stemming
Stemming and Lemmatization both refer to the process of removing all of the affixes (i.e. suffixes, prefixes, etc.) attached to a word in order to keep its lexical base, also known as root or stem or its dictionary form or lemma. The main difference between these two processes is that stemming is usually based on rules that trim word beginnings and endings (and sometimes lead to somewhat weird results), whereas lemmatization makes use of dictionaries and a much more complex morphological analysis.
To provide a more accurate automated analysis of the text, it is important that we remove from play all the words that are very frequent but provide very little semantic information or no meaning at all. These words are also known as stopwords.
There are many different lists of stopwords for every language. However, it is important to understand that you might need to add words to or remove words from those lists depending on the texts you would like to analyze and the analyses you would like to perform.
You might want to do some kind of lexical analysis of the domain your texts come from in order to determine the words that should be added to the stopwords list. Particularly focus on those that might be leading to:
- overfitting of your text classifiers or text extractors to a particular testing set and,
- a lack of generalization of the analyses your classifiers and extractors perform.
Depending on the problem at hand, sequences of numbers, URLs, and some names, for example, might not be relevant for the detection of a topic. If this happens to be the case, you should add those to the stopword list. Look at the example below (from an IT support system):
'Please talk to Professor John Doe (555) 5555 5555 to discuss his computer configuration. He would like to upgrade his HDD and RAM. Thanks.'
Here, the phone number, Professor John Doe, and some words like Please are likely to bear little or no relevance in the detection of the topic of the text: computer configuration. If you add those words to the stopword list, the classifier will not make predictions based on their appearances in texts.
Now that you know the basics of data preparation, let's delve deeper into the fun part: data analysis!
Unstructured text is everywhere and mining data has been brought down to scraping websites, downloading emails, exporting tickets, or the like.
Now, how do you analyze all of this text?
Well, the analysis of unstructured text is not straightforward. There are countless ways of analyzing text to choose from and a great many more you can create yourself. In this section, we will introduce two ways in which you will be able to get insights from your texts, namely, text classification and text extraction.
Text classification (also known as text categorization or text tagging) refers to the process of assigning tags to texts based on its content.
In the past, text classification was done manually, which was time-consuming, inefficient, and inaccurate. At present, we can perform automated text analysis of our data in very little time and get really good results.
Typical text classification tasks include sentiment analysis (i.e. detecting when a text says something positive or negative about a given topic), topic detection (i.e. determining what topics a text talks about), and intent detection (i.e. detecting the purpose or underlying intent of the text), among others, but there are a great many more applications you might be interested in.
In text classification, a rule is essentially a human-made association between a linguistic pattern that can be found in a text and a tag. What rule-based systems do is detecting these handcrafted linguistic patterns in texts and assigning the corresponding tags based on the results of the detections. Usually, rules consist of references to morphological, lexical, or syntactic patterns, but they can also contain references to other components of language, such as semantics or phonology.
Here's an example of a very simple rule for classifying product descriptions according to the type of product described in the text:
(HDD|RAM|SSD|Memory) → Hardware
In this case, the system will assign the Hardware tag to those texts that contain the words HDD, RAM, SSD, or Memory.
The most obvious advantage of rule-based systems is that they are easily understandable by humans. However, creating complex rule-based systems takes a lot of time and a good deal of knowledge of both linguistics and the topics being dealt with in the texts the system is supposed to analyze.
On top of that, rule-based systems are difficult to scale and maintain because adding new rules or modifying the existing ones requires a lot of analysis and testing of the impact of these changes on the results of the predictions.
Machine Learning-based Systems
Machine Learning based systems can make predictions based on what they learn from past observations. In other words, these systems need that you feed them with many examples of texts and the expected predictions (tags) for each of them.
The more consistent and accurate the samples you have fed the classifier with are, the better its predictions will be. That set of tagged texts you feed your system with so that it learns from your texts is called training data.
When you train a machine learning-based classifier, training data has to be transformed into something a machine can understand, that is, vectors (i.e. lists of numbers which encode some information). By using vectors, the system can extract relevant features (pieces of information) which will help it learn from the existing data and make predictions about the texts to come.
There are a number of ways of doing this, but one of the most frequently used is known as the bag of words vectorization. You can learn more about vectorization here.
Once the texts have been transformed into vectors, they are fed into a machine learning algorithm together with their expected output to create a classification model that can choose what features best represent the texts and make predictions about unseen texts:
The trained model will transform unseen text into a vector, extract its relevant features, and make a prediction:
Machine Learning Algorithms
There are many machine learning algorithms used in text classification. The most frequently used are the Naive Bayes family of algorithms (NB), Support Vector Machines (SVM), and deep learning algorithms.
The Naive Bayes family of algorithms is based on Bayes's Theorem and the conditional probabilities of occurrence of the words of a sample text within the words of a set of texts that belong to a given tag. Vectors that represent texts encode information about how likely it is for the words in the text to occur in the texts of a given tag. With this information, the probability of a text's belonging to any given tag in the model can be computed. Once all of the probabilities have been computed for an input text, the classification model will return the tag with the highest probability as the output for that input.
One of the main advantages of this algorithm is results are pretty good when training data is not much.
Support Vector Machines (SVM) is an algorithm that can divide a vector space of tagged texts into two subspaces: one space that contains most of the vectors that belong to a given tag and another subspace that contains most of the vectors that do not belong to that one tag.
Classification models that use SVM at their core will transform texts into vectors and will determine what side of the boundary that divides the vector space for a given tag those vectors belong to. Based on where they land, the model will know if they belong to a given tag or not.
The most important advantage of using SVM is that results are usually better than those obtained if Naive Bayes is used. However, more computational resources are needed in order to use SVM.
Deep Learning is a set of algorithms and techniques inspired by how the human brain works. These algorithms use huge amounts of training data (millions of examples) to generate semantically rich representations of texts which can then be fed into machine learning-based models of different kinds that will make much more accurate predictions than traditional machine learning models:
Hybrid systems usually contain machine learning-based systems at their cores and rule-based systems that are used to further improve the predictions.
Classifier performance is usually evaluated through standard metrics used in the machine learning field. These metrics are accuracy, precision, recall, and F1 score. Understanding what they mean will give you a clearer idea of how good your classifiers are at analyzing your texts.
It is also important to understand that evaluation can be performed over a fixed testing set (i.e. a set of texts for which we know the expected output tags) or by using cross-validation (i.e. a method that splits your training data into different folds so that you can use some subsets of your data for training purposes and some for testing purposes, see below).
In this section, we will introduce you to the standard performance metrics and the cross-validation method so that you can better understand how the performance of your classifiers is evaluated.
Accuracy, Precision, Recall, and F1 score
Accuracy is the number of correct predictions the classifier has made divided by the total number of predictions. In general, accuracy alone is not a good indicator of performance. For example, when categories are imbalanced, that is, when there is one category that contains many more examples than all of the others, predicting all texts as belonging to that category will return high accuracy levels. This is known as the accuracy paradox. To get a better idea of the performance of a classifier, you might want to consider precision and recall instead.
Precision states how many texts were predicted correctly out of the ones that were predicted as belonging to a given tag. In other words, precision takes the number of texts that were correctly predicted as positive for a given tag and divides it by the number of texts that were predicted (correctly and incorrectly) as belonging to the tag.
We have to bear in mind that precision only gives information about the cases where the classifier predicts that the text belongs to a given tag. This might be particularly important, for example, if you would like to generate automated responses for user messages. In this case, before you send an automated response you want to know for sure you will be sending the right response, right? In other words, if your classifier says the user message belongs to a certain type of messages, you would like the classifier to make the right guess. This means you would like a high precision for that type of messages.
Recall states how many texts were predicted correctly out of the ones that should have been predicted as belonging to a given tag. In other words, recall takes the number of texts that were correctly predicted as positive for a given tag and divides it by the number of texts that were either predicted correctly as belonging to the tag or that were incorrectly predicted as not belonging to the tag.
Recall might prove useful when routing support tickets to the appropriate team, for example. It might be desired for an automated system to detect as many tickets as possible for a critical tag (for example tickets about 'Outrages / Downtime') at the expense of making some incorrect predictions along the way. In this case, making a prediction will help perform the initial routing and solve most of these critical issues ASAP. If the prediction is incorrect, the ticket will get rerouted by a member of the team. When processing thousands of tickets per week, high recall (with good levels of precision as well, of course) can save support teams a good deal of time and enable them to solve critical issues faster.
The F1 score is the harmonic means of precision and recall. It tells you how well your classifier performs if equal importance is given to precision and recall. In general, F1 score is a much better indicator of classifier performance than accuracy is.
Cross-validation is quite frequently used to evaluate the performance of text classifiers. The method is simple. First of all, the training dataset is randomly split into a number of equal-length subsets (e.g. 4 subsets with 25% of the original data each). Then, all the subsets except for one are used to train a classifier (in this case, 3 subsets with 75% of the original data) and this classifier is used to predict the texts in the remaining subset. Next, all the performance metrics are computed (i.e. accuracy, precision, recall, f1, etc.). Finally, the process is repeated with a new testing fold until all the folds have been used for testing purposes.
Once all folds have been used, the average performance metrics are computed and the evaluation process is finished.
Text Extraction refers to the process of recognizing structured pieces of information from unstructured text.
For example, it might be useful to automatically detect the most relevant keywords from a piece of text, identify names of companies in a news article, detect lessors and lessees in a financial contract, or identify prices on product descriptions. Just like text classification, text extraction can be performed automatically or manually, with the latter being tremendously more time-consuming and inefficient.
There are different ways in which automated text extraction can be implemented. In this section, we introduce some approaches which are widely accepted and return very good results.
Regular Expressions (a.k.a. regexes) work as the equivalent of the rules defined in classification tasks. In this case, a regular expression defines a pattern of characters that will be associated with a tag.
For example, the pattern below will detect most email addresses in a text if they preceded and followed by spaces:
By detecting this match in texts and assigning it the email tag, we can create a rudimentary -but functional- email address extractor.
There are obvious pros and cons of this approach. On the plus side, you can create text extractors quickly and the results obtained can be good, provided you can find the right patterns for the type of information you would like to detect. On the minus side, regular expressions can get extremely complex and might be really difficult to maintain and scale, particularly when many expressions are needed in order to extract the desired patterns.
Conditional Random Fields
Conditional Random Fields (CRF) is a statistical approach often used in machine-learning-based text extraction. This approach learns the patterns to be extracted by weighing a set of features of the sequences of words that appear in a text. Through the use of CRFs, we can add multiple variables which depend on each other to the patterns we use to detect information in texts, such as syntactic or semantic information.
This usually generates much richer and complex patterns than using regular expressions and can potentially encode much more information. However, more computational resources are needed in order to implement it since all the features have to be calculated for all the sequences to be considered and all of the weights assigned to those features have to be learned before determining whether a sequence should belong to a tag or not.
One of the main advantages of the CRF approach is its generalization capacity. Once an extractor has been trained using the CRF approach over texts of a specific domain, it will have the ability to generalize what it has learned to other domains reasonably well.
Extractors are sometimes evaluated by calculating the same standard performance metrics we have explained above for text classification, namely, accuracy, precision, recall, and f1 score. However, these metrics do not account for partial matches of patterns. In order for an extracted segment to be a true positive for a tag, it has to be a perfect match with the segment that was supposed to be extracted.
Consider the following example:
'Your flight will depart on June 13, 2019 at 03:30 PM from SFO.'
If we created a date extractor, we would expect it to return June 13, 2019 as a date from the text above, right? So, if the output of the extractor were June 13, 2019, we would count it as a true positive for the tag DATE.
But, what if the output of the extractor were June 13? Would you say the extraction was bad? Would you say it was a false positive for the tag DATE? To capture partial matches like this one, some other performance metrics can be used to evaluate the performance of extractors. One example of this is the ROUGE family of metrics.
ROUGE (Recall-Oriented Understudy for Gisting Evaluation) is a family of metrics used in the fields of machine translation and automatic summarization that can also be used to assess the performance of text extractors. These metrics basically compute the lengths and number of sequences that overlap between the source text (in this case, our original text) and the translated or summarized text (in this case, our extraction).
Depending on the length of the units whose overlap you would like to compare, you can define ROUGE-n metrics (for units of length n) or you can define the ROUGE-LCS or ROUGE-L metric if you intend to compare the longest common sequence (LCS).
So, you're getting up to speed with automatic text analysis and now you're able to process complex data in a fast and effective way using AI. What's next? The answer is simple: data visualization.
When you use text analysis with machine learning to classify or extract specific text data, the outcome is either:
- An API response in JSON format (if you are a coder).
- A CSV or an Excel file (if you can't code).
But, what can you do with these results?
Let's imagine you've automatically analyzed 10,000 responses from open-ended questions in a customer feedback survey, and now you're going to present the results within your company. What if you could transform the raw results into well-designed data graphics and display the content in a visual, impactful and engaging way?
Or, say you have to analyze thousands of product reviews. Wouldn't be it easier to just look at charts and graphs to detect trends and find valuable insights for your reports, instead of staring at a boring Excel spreadsheet?
Data visualization * boosts the value* of the results obtained with text mining. By using business intelligence tools, you can transform complex concepts into compelling and simple information. Data graphics also make it easier to establish relationships and observe patterns. It's all about high-quality insights that lead to smart data-oriented business decisions!
There are several data visualization tools that you can use to process text analysis results. Let's take a closer look at the three most popular:
- Google Data Studio
This is Google's free and easy-to-use visualization tool that allows you to create interactive reports using a wide variety of data. To get started, you'll need to connect your data to the platform (you can connect to more than 100 different sources!).
Once you've imported your data you can use different tools to design your report and turn your data into an attractive story. Finally, you can share the results with individuals or teams, publish them on the Web or embed them on your website.
Looker is a business data analytics platform designed to direct meaningful data to everyone in a company. The idea is to allow teams to have a 'bigger picture' about what's happening in their company.
This platform connects to different databases and automatically creates a data model, which can be fully customized to meet specific needs. You can use data to build your own friendly dashboards and reports, interact with data in real-time and share it with your team members.
Here are some instructions to get started.
Tableau is a business intelligence and data visualization tool, which makes it easy to work with data and create intuitive visual analytics. With a very user-friendly approach (there are no technical skills required), Tableau allows organizations to work with almost any existing data source and provides powerful visualization options.
The Tableau suite offers different products; some of them are developer tools, while others are sharing tools for non-coders. There's a trial version available for anyone wanting to give it a go. Check out this video if you want to learn how to get started with Tableau.
Businesses generate lots of information every day and most of it is unstructured text data. Whether you work in sales, customer support, marketing or product, analyzing raw text data is extremely important. When done manually, the tagging process is time-consuming, repetitive and inaccurate. But above all, it leaves employees with less time to invest in other tasks, the ones where you can really make a difference and set the bar high for a successful customer experience.
Artificial intelligence, text analysis, machine learning: all these terms seem too complex, broad and difficult to grasp. But what if you could benefit from these technologies without needing to actually invest a lot of time in learning or building tools from scratch? What if you didn't need to be a developer or a machine learning expert to get started with text analysis?
Now that you've read this section, you probably have a few ideas of data sources you can use for automated text analysis. Also, you already know what a text classifier and a text extractor are, and how they can be trained with machine learning. Finally, you've learned how data visualization tools can help you to make your results more appealing.
So what's next? The following section will go into depth about the different applications of text analysis and the impact this can have on your business. We'll provide case studies and examples to explain how you can use text analysis in a wide number of areas within your company, like social media monitoring, brand monitoring, customer support, voice of customer (VoC), business intelligence, sales marketing, product analytics, and knowledge management.
Use Cases and Applications
Did you know that 80 percent of business data is text? Text is present in every major business process, from support tickets to product feedback and customer interactions. But analyzing heaps of text can be quite daunting. That's why text analysis is growing in popularity, helping more companies to automate different tasks and processes with machine learning.
Text analysis has a broad range of business applications and use cases. Some businesses use this technology to maximize efficiency and reduce the time employees spend doing repetitive tasks that can potentially have a high turnover impact. Others are hoping to better understand customer insights without having to sort through millions of social media posts, online reviews, and survey responses.
If you work in customer experience, product, marketing or sales, there are quite a few text analysis applications that can help you automate manual processes and get better insights that won't require you to be tech savvy. In this section, we'll cover the following use cases and applications:
- Social Media Monitoring
- Brand Monitoring
- Customer Service
- Voice of Customer & Customer Feedback
- Business Intelligence
- Sales and Marketing
- Product Analytics
- Knowledge Management
Let's dive in!
Social Media Monitoring
Let's say you work for Uber and you want to know what users are saying about the brand. You've read some positive and negative feedback on Twitter and Facebook. But 500 million tweets are sent each day, and Uber has thousands of mentions on social media every month. Can you imagine analyzing all of them manually? Where do you start? All this raw text data needs to be converted into numbers. This is where text analysis with machine learning plays a crucial role.
A useful option is sentiment analysis, which analyzes the opinion about a given subject within a text. By analyzing your social media mentions with a sentiment analysis model, you can automatically categorize them into Positive, Neutral or Negative. If you also analyze these mentions with a topic classifier, you can also understand what they are talking about. By running aspect-based sentiment analysis, you can automatically pinpoint the reasons behind positive or negative mentions and get insights such as:
- The top complaint about Uber on Social Media?
- The success rate of Uber's customer service - are people happy or are annoyed with it?
- What Uber users like about the service when they mention Uber in a positive way?
Now, let's say you've just added a new service to Uber. For example, Uber Eats. It's a crucial moment, and your company wants to know what people are saying about Uber Eats so that you can fix any glitches as soon as possible, and polish the best features. You can also use aspect-based sentiment analysis on your Facebook, Instagram and Twitter profiles for any Uber Eats mentions and discover things such as:
- Are people happy with Uber Eats so far?
- What is the most urgent issue to fix?
- How can we incorporate positive stories into our marketing and PR communication?
Not only can you use text analysis to keep tabs on your brand's social media mentions, but you can also use it to monitor your competitors' mentions as well. Is a client complaining about a competitor's service? That gives you a chance to attract potential customers and show them how much better your brand is.
Have you ever had to deal with a barrage of negative comments on the internet? It's quite a stressful task, to say the least, so spotting them on social media, forums, blogs and review sites as soon as possible could reduce the negative impact on your brand.
The power of negative reviews is quite strong: 40% of consumers are put off from buying if a business has negative reviews. An angry customer complaining about poor customer service can spread like wildfire within minutes: a friend shares it, then another, then another… And before you know it, the negative comments have gone viral.
So, where do you start? First, you can do some web scraping to automatically collect reviews or news articles. Here are different tools to get this done:
The next step will be running text analysis models on the scraped data to get useful insights about online conversations that mention your brand. This can help you do all of this:
- Understand how your brand reputation evolves over time.
- Compare your brand reputation to your competitor's.
- Identify which aspects are damaging your reputation.
- Pinpoint which elements are boosting your brand reputation on online media.
- Identify potential PR crises so you can deal with them ASAP.
- Tune into data from a specific moment, like the day of a new product launch or IPO filing. Just run a sentiment analysis on social media and press mentions on that day, to find out what people said about your brand.
Despite many people's fears and expectations, text analysis doesn't mean that customer service will be entirely machine-powered. It just means that businesses will have more seamless processes so that teams can spend more time-solving problems that require human interaction. That way businesses will be able to increase retention, given that 89 percent of customers change brands because of poor customer service. But, how can text analysis, assist your company's customer service?
Ticket Tagging: Make the Most out of Ticket Tagging
When a customer complaint, comment, or request arises, customer service representatives have to take the time to categorize each ticket before addressing it. It's a fairly simple task, but it's a boring and time-consuming process. Each tag takes at least two clicks to add. And even if a support agent takes the time to read each ticket and choose the most appropriate tags, they're not always consistent.
That's why we need to start delegating the task to machines. Text analysis automatically identifies topics, and tags each ticket. With a bit of previous training, yes, but without customer service agents becoming frustrated. Here's how it works :
- The model analyzes the language and expressions a customer uses, for example, 'I didn't get the right order'.
- Then, it compares it to other similar conversations.
- Finally, it finds a match and tags the ticket automatically. In this case, it could be under a Shipping Problems tag.
This process repeats itself every time a new ticket comes in, freeing customer agents to focus on more important tasks.
Ticket Routing & Triage: Find the Right Person for the Job
With text analysis, team members can save time they'd have otherwise spent reading tickets and manually assigning them to the qualified rep. This can now be done automatically and in real-time with a text analysis model that can pinpoint what each ticket is about, and then route it accordingly to the most appropriate person, whether it's because of their skills, appointed tasks, or native language.
For example, for a SaaS company that receives a customer ticket asking for a refund, the text mining system will identify which team usually handles billing issues and send the ticket to them. If a ticket says something like 'How can I integrate your API with python?', it would go straight to the team in charge of helping with Integrations.
Ticket Analytics: Learn More from your Customers
What is commonly assessed to determine the performance of a customer service team? Common KPIs are first response time, average time to resolution (i.e. how long it takes your team to resolve issues), and customer satisfaction (CSAT). And, let's face it, overall client satisfaction has a lot to do with the first two metrics.
But how do we get actual CSAT insights from customer conversations? How can we identify if a customer is happy with the way an issue was solved? Or if they have expressed frustration with the handling of the issue?
In this situation, aspect-based sentiment analysis could be used. This type of text analysis delves into the feelings and topics behind the words on different support channels, such as support tickets, chat conversations, emails and CSAT surveys. A text analysis model can understand words or expressions to define the support interaction as Positive, Negative, or Neutral, understand what was mentioned (e.g. Service or UI/UX) and even determine the sentiments behind the words (e.g. Sadness, Anger, etc.).
Urgency Detection: Prioritize Urgent Tickets
“Where do I start?” is a question most customer service representatives often ask themselves. Urgency is definitely a good starting point, but how do we define the level of urgency without wasting valuable time deliberating?
A text mining model can define the urgency level of a customer ticket and tag it accordingly. Support tickets with words and expressions that denote urgency, such as 'as soon as possible' or 'right away', are duly tagged as Priority.
To see how text analysis works to detect urgency, check out this MonkeyLearn urgency detection demo model.
Voice of Customer & Customer Feedback
Once you get a customer, retention is key, since acquiring new clients is five to 25 times more expensive than retaining the ones you already have. That's why paying close attention to the voice of the customer can give your company a clear picture of the level of client satisfaction and, consequently, of client retention. Also, it can give you actionable insights to prioritize the product roadmap from a customer's perspective.
The thing is that most of your VoC is probably written text, so how can you gain knowledge from thousands of feedback entries?
Analyzing NPS Responses
Maybe your brand already has a customer satisfaction survey in place, the most common one being the Net Promoter Score (NPS). This survey asks the question 'How likely is it that you would recommend [brand] to a friend or colleague?'. The answer is a score from 0-10 and the result is divided into three groups: the promoters, the passives, and the detractors.
But here comes the tricky part: there's an open-ended follow-up question at the end 'Why did you choose X score?' The answer can provide your company with invaluable insights. Without the text, you're left guessing what went wrong. The problem is going through these open-ended responses is hard:
- Customer reps have to sort through and categorize every feedback entry manually. This means that hours are lost solely on reading.
They have to tag it in a consistent way, which is hard for one employee, let alone a whole team.
- To solve these problems and get the evidence to surface, text analysis can be applied in different ways.
You can do what Promoter.io did: extract the main keywords of your customers' feedback to understand what's being praised or criticized about your product. Is the keyword 'Product' mentioned mostly by promoters or detractors? With this info, you'll be able to use your time to get the most out of NPS responses and start taking action.
Another option is following Retently's footsteps using text analysis to classify your feedback into different topics, such as 'Customer Support', 'Product Design', and 'Product Features', among others. After, they analyzed each tag with sentiment analysis to see how positively or negatively clients feel about each topic. Now they know they're on the right track with product design, but still have to work on product features.
Analyzing Customer Surveys
Does your company have another customer survey system? If it's a scoring system or closed-ended questions, it'll be a piece of cake to analyze the responses: just crunch the numbers.
However, if you have an open-text survey, whether it's provided via email or it's an online form, you can stop manually tagging every single response by letting text analysis do the job for you. Besides saving time, you can also have consistent tagging criteria without errors, 24/7.
Data analysis is at the core of every business intelligence operation. Now, what can a company do to understand, for instance, sales trends and performance over time? With numeric data, a BI team can identify what's happening (such as sales of X are decreasing) – but not why. Numbers are easy to analyze, but they are also somewhat limited. Text data, on the other hand, is the most widespread format of business information and can provide your organization with valuable insight into your operations. The thing is… analyzing text manually is frustrating 😫!
Let's say you work for a SaaS startup. You're part of a small but mighty team working on a great new product that solves a real problem. After months of hard work, you successfully launch your product, the first sales are made, and the business grows. Your customers are leaving positive feedback, the team starts expanding, and your startup experiences the so-called hockey-stick growth. Then, after a couple of high-growth years, somehow, business starts slowing down. The flow of sales prospects is still good, but you start to have high customer churn. It's harder and harder for your sales reps to close new deals and the future growth prospects begin to stall. What changed?
Your team can lean on text analysis to find out what is going on with your company, offering concrete insights for key decision-making.
For example, you can run keyword extraction and sentiment analysis on your social media mentions to understand what people are complaining about regarding your brand.
You can also run aspect-based sentiment analysis on customer reviews that mention poor customer experiences. After all, 67% of consumers list bad customer experience as one of the primary reasons for churning. Maybe it's bad support, a faulty feature, unexpected downtime, or a sudden price change. Analyzing customer feedback can shed a light on the details, and the team can take action accordingly.
But let's not limit this to internal data… What about the competitors? What are their reviews saying? Run them through your text analysis model and see what they're doing right and wrong and improve your decision-making.
Sales and Marketing
Prospecting is the most difficult part of the sales process. And it's getting harder and harder. The sales team always want to close deals, which requires making the sales process more efficient. But 27% of sales agents are spending over an hour a day on data entry work instead of selling, meaning critical time is lost to administrative work and not closing deals.
Text analysis takes the heavy lifting out of manual sales tasks, including:
- Updating the deal status as 'Not interested' in your CRM.
- Qualifying your leads based upon company descriptions.
- Identifying leads on social media that express buying intent.
GlassDollar, a company that links founders to potential investors, is using text analysis to find the best quality matches. How? They use text analysis to classify companies using their company descriptions. The results? They saved themselves days of manual work, and predictions were 90% accurate after training a text classification model. You can learn more about their experience with MonkeyLearn here.
Not only can text analysis automate manual and tedious tasks, but it can also improve your analytics to make the sales and marketing funnels more efficient. For example, you can automatically analyze the responses from your sales emails and conversations to understand, let's say, a drop in sales:
- What are the blocks to completing a deal?
- What sparks a customer's interest?
- What are customer concerns?
Now, Imagine that your sales team's goal is to target a new segment for your SaaS: people over 40. The first impression is that they don't like the product, but why? Just filter through that age group's sales conversations and run them on your text analysis model. Sales teams could make better decisions using in-depth text analysis on customer conversations.
Finally, you can use machine learning and text analysis to provide a better experience overall within your sales process. For example, Drift, a marketing conversational platform, integrated MonkeyLearn API to automatically allow recipients to opt out of sales emails based on how they reply.
It's time to boost sales and stop wasting valuable time with leads that don't go anywhere. Xeneta, a sea freight company, developed a machine learning algorithm and trained it to identify which companies were potential customers, based on the company descriptions gathered through FullContact (a SaaS company that has descriptions of millions of companies).
You can do the same or target users that visit your website to:
- Get information about where potential customers work using a service like Clearbit and classify the company according to its type of business to see if it's a possible lead.
- Extract information to easily learn the user's job position, the company they work for, its type of business and other relevant information.
- Hone in on the most qualified leads and save time actually looking for them: sales reps will receive the information automatically and start targeting the potential customers right away.
We've covered the importance of text analysis applications for customer feedback, but we haven't delved into the impact this machine learning tool can have on a brand's product analytics. For argument's sake, let's imagine your startup has an app on the Google Play store. You're receiving some unusually negative comments. What's going on?
It can be daunting to locate and tackle a problem if you first have to read through a thousand reviews. Alternatively, you can discover what's going on within minutes by using a text analysis model that groups reviews into different tags like 'Ease of Use' and 'Integrations'. Then, you can run them on the sentiment analysis model to find out whether customers are talking about products positively or negatively. Finally, graphs and reports can be created to visualize and prioritize product problems. We did this with reviews for Slack from the product review site Capterra and got some pretty interesting insights. Here's how:
We analyzed reviews with aspect-based sentiment analysis and categorized them into main topics and sentiment.
We extracted keywords with the keyword extractor to get some insights into why reviews that are tagged under 'Performance-Quality-Reliability' tend to be negative.
A quick example: your social media app targets young people perfectly, reviews are mostly positive, but you want to expand and also target an older demographic. With aspect-based sentiment analysis, your product team may learn that many reviews under the 'Ease of Use' tag shed light on this demographic: some features might need to be tailored to non-digital natives. But which features? By running the negative reviews on a keyword extractor you'll track them down within seconds.
CEOs value the practice of sharing knowledge but often find that it's a struggle to encourage teams to upload information to the knowledge management system in place (if there is one)... or just simply to the Google Drive. After all, communication is key to any relationship. Knowledge management requires teams to manually upload text-form data to a platform or database, and with text analysis, this task can be made automatically.
For example, HR is tasked with reading through and classifying CVs. If it's a big company, the HR department probably receives hundreds of CVs per day. Whether the company is looking to fill a new job position or not, wouldn't it be helpful to have resumés classified by background, area or qualifications? Plus, reading through CVs takes up valuable time, so a promising candidate may get lost amid all the mediocre resumés. To make things easier, HR could:
- Apply a text classifier that automatically tags CVs to find candidates more easily.
- Use a text extractor to identify and extract main aspects of the CV, including 'Full Name', 'Degree', 'Job Experience', and 'Skills'.
Of course, text analysis tools can be used for other purposes, like extracting keywords or creating a summary of long documents, such as contracts, so that they can be read in a few minutes to get a general idea of the topic.
We've all heard the saying that ‘knowledge is power', and we couldn't agree more. Unstructured text data can be transformed into useful data that gives a business knowledge about how to boost its sales, reduce turnover or even understand its clients. And all this is made possible thanks to text analysis.
Text analysis is the process of obtaining structured knowledge from natural language text. It's the foundation of Natural Language Processing (NLP) and as such, if you're building an NLP solution, you'll inevitably have to learn text analysis.
Luckily, it's easy to get started with text analysis: there are a lot of useful resources available that can help you get your feet wet. In this section, we'll cover the following resources:
Let's get started!
Text Analysis APIs
Open Source Libraries
If a commercial off-the-shelf solution for your text analysis needs is not available, you can build your own using widely-available open source solutions.
Python is the most widely-used language in scientific computing, period. Tools like NumPy and SciPy have established it as a fast, dynamic language that calls C and Fortran libraries where performance is needed.
These things, combined with a thriving community and a diverse set of libraries to implement NLP models has made Python one of the most preferred programming languages for doing text analysis.
NLTK, the Natural Language Toolkit, is a best-of-class library for text analysis tasks. NLTK is used in many university courses, so there's plenty of code written with it and no shortage of users familiar with both the library and the theory of NLP who can help answer your questions.
Unlike NLTK, which is a research library, SpaCy aims to be a battle-tested, production-grade library for text analysis.
Scikit-learn is a complete and mature machine learning toolkit for Python built on top of NumPy, SciPy, and matplotlib, which gives it stellar performance and flexibility for building text analysis models.
Developed by Google, TensorFlow by far is the most widely used library for distributed deep learning. Looking at this graph we can see that TensorFlow is ahead of the competition:
This has effects further down the line: most libraries, ready-made models, and notebooks are built with Tensorflow in mind. So, knowing how to use TensorFlow is important if you're interested in getting into machine learning and text analysis.
PyTorch is a deep learning platform built by Facebook and aimed specifically at deep learning. PyTorch is a Python-centric library, which allows you to define much of your neural network architecture in terms of Python code, and only internally deals with lower-level high-performance code.
Keras is a widely-used deep learning library written in Python.
It's designed to enable rapid iteration and experimentation with deep neural networks, and as a Python library, it's uniquely user-friendly.
An important feature of Keras is that it provides what is essentially an abstract interface to deep neural networks. The actual networks can run on top of Tensorflow, Theano, or other backends. This backend independence makes Keras an attractive option in terms of its long-term viability.
The permissive MIT license makes it attractive to businesses looking to develop proprietary models.
R is the pre-eminent language for any statistical task. Its collection of libraries (13,711 at the time of writing on CRAN far surpasses any other programming language capabilities for statistical computing and is larger than many other ecosystems. In short, if you choose to use R for anything statistics-related, you won't find yourself in a situation where you have to reinvent the wheel, let alone the whole stack.
caret is an R package designed to build complete machine learning pipelines, with tools for everything from data ingestion and preprocessing, feature selection, and tuning your model automatically.
The Machine Learning in R project (mlr for short) provides a complete machine learning toolkit for the R programming language that's frequently used for text analysis.
Java needs no introduction. The language boasts an impressive ecosystem that stretches beyond Java itself and includes the libraries of other The JVM languages such as The Scala and Clojure. Beyond that, the JVM is battle-tested and has had thousands of person-years of development and performance tuning, so Java is likely to give you best-of-class performance for all your text analysis NLP work.
Stanford's CoreNLP project provides a battle-tested, actively maintained NLP toolkit. While it's written in Java, it has APIs for all major languages, including Python, R, and Go.
The Apache OpenNLP project is another machine learning toolkit for NLP. It can be used from any language on the JVM platform.
Weka is a GPL-licensed Java library for machine learning, developed at the University of Waikato in New Zealand. In addition to a comprehensive collection of machine learning APIs, Weka has a graphical user interface called the Explorer, which allows users to interactively develop and study their models.
Weka supports extracting data from SQL databases directly, as well as deep learning through the deeplearning4j framework.
Using a SaaS API for text analysis has a lot of advantages:
- No setup:
A tool implies the rest of its ecosystem. Getting started with a library for Python, for example, means setting up a Python installation and getting the Python package manager running.
As Neal Stephenson remarked, "every broken python installation is broken in its own way". Getting all the pieces working can be a painful process, and doubly so if you're working on Windows.
And that's just to get the programming language and its tooling working. Libraries might involve specific set up steps, like installing system libraries for linear algebra, getting GPU drivers, etc.
- No code:
SaaS APIs provide ready to use solutions. You give them data and they return the analysis after some time period. Every other concern -- performance, scalability, logging, architecture, tools, etc. -- is offloaded to the party responsible for maintaining the API.
The only development effort requires is when it comes to integrating with your codebase: you just need to write code to call the API and get the results back.
SaaS APIs for text analysis usually provide ready-made client libraries for a number of programming languages, which simplifies development effort further: instead of writing a client for a REST API, all you have to do is write a client for another library.
- Easy Integration:
SaaS APIs usually provide ready-made integrations for tools like Zapier or Google Sheets. This will allow you to build a truly no-code solution: you can arrange integrations to feed data from a cloud source directly into the SaaS API and store the results back in your cloud. All of this with zero lines of code and without the need to have a programming background.
Some of the most well-known SaaS solutions and APIs for text analysis include:
- Google Cloud NLP
- IBM Watson
- Amazon Comprehend
If you talk to any data science professional, they'll tell you that the true bottleneck to building better models is not new and better algorithms, but more data.
Indeed, in machine learning, data is king: a simple model, given tons of data, is likely to outperform one that uses every trick in the book to turn every bit of training data into a meaningful response.
So, here are some high-quality datasets you can use to get started:
Reuters news dataset: one the most popular datasets for text classification; it has thousands of articles from Reuters tagged with 135 categories according to their topics, such as Politics, Economics, Sports, and Business.
20 Newsgroups: a very well-known dataset that has more than 20k documents across 20 different topics.
Product reviews: a pretty big dataset with millions of customer reviews from products on Amazon. Besides the reviews, it provides useful metadata such as star ratings, super useful for training a sentiment analysis model.
Twitter airline sentiment on Kaggle: another widely used dataset for getting started with sentiment analysis. It contains more than 15k tweets about airlines (tagged as positive, neutral, or negative).
First GOP Debate Twitter Sentiment: another useful dataset with more than 14,000 labeled tweets (positive, neutral, and negative) from the first GOP debate in 2016.
Other Popular Datasets
Spambase: this dataset contains 4,601 emails tagged as spam and not spam.
SMS Spam Collection: another dataset for spam detection. It has more than 5k SMS messages tagged as spam and not spam.
Hate speech and offensive language: a dataset with more than 24k tagged tweets grouped into three tags: clean, hate speech, and offensive language.
Finding high-volume and high-quality training datasets are the most important part of text analysis, more important than the choice of the programming language or tools for creating the models. Remember, the best-architected machine-learning pipeline is worthless if its models are backed by unsound data.
Text Analysis Tutorials
The best way to learn is by doing. Theory comes later.
First, we'll go through programming-language-specific tutorials using open-source tools for text analysis. These will help you deepen your understanding of the available tools for your platform of choice.
Then, we'll take a step-by-step tutorial of MonkeyLearn so you can get started with text analysis right away.
Tutorials Using Open Source Libraries
In this section, we'll look at various tutorials for text analysis in the main programming languages for machine learning that we listed above.
Other useful tutorials include:
WordNet with NLTK: Finding Synonyms for words in Python: this tutorial shows you how to build a thesaurus using Python and WordNet.
Tokenizing Words and Sentences with NLTK: this tutorial shows you how to use NLTK's language models to tokenize words and sentences.
spaCy 101: Everything you need to know: part of the official documentation, this tutorial shows you everything you need to know to get started using SpaCy.
This tutorial shows you how to build a WordNet pipeline with SpaCy.
Furthermore, there's the official API documentation, which explains the architecture and API of SpaCy.
If you prefer long-form text, there are a number of books about or featuring SpaCy:
- Introduction to Machine Learning with Python: A Guide for Data Scientists.
- Practical Machine Learning with Python.
- Text Analytics with Python.
The official scikit-learn documentation contains a number of tutorials on the basic usage of scikit-learn, building pipelines, and evaluating estimators.
Scikit-learn Tutorial: Machine Learning in Python shows you how to use scikit-learn and Pandas to explore a dataset, visualize it, and train a model.
For readers who prefer books, there are a couple of choices:
Our very own Raúl Garreta wrote this book: Learning scikit-learn: Machine Learning in Python.
Additionally, the book Hands-On Machine Learning with Scikit-Learn and TensorFlow introduces the use of scikit-learn in a deep learning context.
The official Keras website has extensive API as well as tutorial documentation. For readers who prefer long-form text, the Deep Learning with Keras book is the go-to resource. The book uses real-world examples to give you a strong grasp of Keras.
Practical Text Classification With Python and Keras: this tutorial implements a sentiment analysis model using Keras, and teaches you how to train, evaluate, and improve that model.
Text Classification in Keras: this article builds a simple text classifier on the Reuters news dataset. It classifies the text of an article into a number of categories such as sports, entertainment, and technology.
TensorFlow Tutorial For Beginners introduces the mathematics behind TensorFlow and includes code examples that run in the browser, ideal for exploration and learning. The goal of the tutorial is to classify street signs.
The book Hands-On Machine Learning with Scikit-Learn and TensorFlow helps you build an intuitive understanding of machine learning using TensorFlow and scikit-learn.
Finally, there's the official Get Started with TensorFlow guide.
The official Get Started Guide from PyTorch shows you the basics of PyTorch. If interested in something more practical, check out this chatbot tutorial; it shows you how to build a chatbot using PyTorch.
The Deep Learning for NLP with PyTorch tutorial is a gentle introduction to the ideas behind deep learning and how they are applied in PyTorch.
Finally, the official API reference explains the functioning of each individual component.
A Short Introduction to the Caret Package shows you how to train and visualize a simple model. A Practical Guide to Machine Learning in R shows you how to prepare data, build and train a model, and evaluate its results. Finally, you have the official documentation is super useful to get started with Caret.
For those who prefer long-form text, on arXiv we can find an extensive mlr tutorial paper. This is closer to a book than a paper and has extensive and thorough code samples for using mlr. There's also the official mlr cheatsheet, a handy resource to have when debugging.
If interested in learning about CoreNLP, you should check out Linguisticsweb.org's tutorial which explains how to quickly get started and perform a number of simple NLP tasks from the command line. Moreover, this CloudAcademy tutorial shows you how to use CoreNLP and visualize its results. You can also check out this tutorial specifically about sentiment analysis with CoreNLP. Finally, there's this tutorial on using CoreNLP with Python that is useful to get started with this framework.
First things first: the official Apache OpenNLP Manual should be the starting point. The book Taming Text was written by an OpenNLP developer and uses the framework to show the reader how to implement text analysis. Moreover, this tutorial takes you on a complete tour of OpenNLP, including tokenization, part of speech tagging, parsing sentences, and chunking.
The Weka library has an official book Data Mining: Practical Machine Learning Tools and Techniques that comes handy for getting your feet wet with Weka.
If you prefer videos to text, there are also a number of MOOCs using Weka:
Data Mining with Weka: this is an introductory course to Weka.
More Data Mining with Weka: this course involves larger datasets and a more complete text analysis workflow.
Advanced Data Mining with Weka: this course focuses on packages that extend Weka's functionality.
The Text Mining in WEKA Cookbook provides text-mining-specific instructions for using Weka.
How to Run Your First Classifier in Weka: shows you how to install Weka, run it, run a classifier on a sample dataset, and visualize its results.
Practical Tutorial with MonkeyLearn
Let's imagine that you're concerned about the productivity of your customer support team. They tell you that they lose hours every day sorting through incoming tickets. Manually classifying topics and sending each ticket to the appropriate department is making their job very difficult.
You need a solution that will help your customer support team sort through support tickets, and you'd like to know how machine learning can do the heavy lifting. Your dev team just doesn't have the time or experience in machine learning, and you don't want to hire new staff.
The good news is that MonkeyLearn offers a simple and easy-to-use platform to build your own text analysis models, without needing to code or having prior machine learning knowledge or experience.
But how does MonkeyLearn work? Here's a short step-by-step tutorial on how to build and train two different text analysis models: a text classifier and a text extractor.
Let's start with the classifier!
Create Your Own Text Classifier
In this six-step tutorial, we'll be able to cover the basics on how to create a text classifier model and put it to the test. Let's get going.
1. Create a new model
2. Now choose 'Topic Classification'
3. Upload your data for training the classifier
The first thing your new model needs is data. You can import data from CSV or Excel files, and from Twitter, Gmail, Zendesk, Freshdesk and other third-party integrations offered by MonkeyLearn:
Once you've uploaded your data, you'll start training your model.
4. Define tags
The next step is to determine the tags you want your text classifier to use when sorting data:
5. Start training your model by tagging examples
For your text classifier to understand your criteria, and to automatically tag the data you feed it, you'll need to manually input some examples. Around four samples per tag should be enough for your model to have a very basic understanding of the information it needs to classify. Nevertheless, the more time you spend training it, the more accurate your model will be.
Training your model is super easy. Just assign the appropriate tag to each piece of text, like in the example below:
If you start noticing that some examples are already tagged, that's machine learning getting the job done! Remember that these are the model's first steps with your data, so it may make some mistakes along the way. Just retag if something's off and you'll see more accurate results in no time.
6. Test your classifier
The last step is to test your text classifier. Once you have finished tagging the initial set of examples, head over to the 'Run' tab and add some more samples into the text box for your model to analyze and make predictions:
If you still see mistakes, you can always return to the 'Build' tab for further training.
Put Your New Text Classifier To Work
Now that your new topic classifier is up and running, all you need to do is upload new data and let the model do its thing. With MonkeyLearn you can simply upload more CSV and Excel files to analyze new data or use one of our integrations:
Create Your Own Text Extractor
So now onto text extraction. Let's say that the results from your latest customer satisfaction survey have just arrived and you need to find a way of sifting through open answers to replace time-wasting manual procedures. These answers contain important information about how customers view your service. Your aim is to extract the most relevant keywords from these answers and gain some useful insights.
Via MonkeyLearn you can leave all this work to machine learning models, allowing you to quickly and simply extract keywords and expressions from the text provided by your customers. The result? A quick and effective way of finding out which are the most common topics your customers talk about when asked about your services.
Mix this with a sentiment analysis check and you'll also know how they feel towards those topics, for example: positively, negatively or neutrally.
So, let's dive into a step-by-step tutorial that'll teach you how to build your own text extractor with MonkeyLearn. You'll find a quick rundown on how to easily set up your first text extraction model, train it and test it in just a couple of minutes.
1. Create a new extractor
2. Select how you want to feed the model data
Now you can upload data from platforms like Twitter or Gmail, or customer support services like Front, Zendesk and Freshdesk, as well as RSS and either CSV or Excel files. Once you've done this, the system will automatically import your data and allow you to begin training your new text extraction model:
3. Create tags for your model to extract
You just need a minimum of two tags to start training your text extraction model, but you can always add more along the way if you feel that there's important data missing:
In this example above, we are choosing tags that will give us insights into how customers view three important aspects of our service. We've extracted data from our latest satisfaction survey using the tags Support, UX/UI, and Pricing.
4. Train your new text extractor
Once you've told your extractor what tags to look out for, it's time for you to manually tag some examples by highlighting words and expressions related to them. After just a little training, you'll start to notice that the extractor begins predicting tags based on your criteria:
It's always a good idea to go back to the 'Build' tab and continue training your model so that it's able to detect language variations you want it to recognize.
Put Your New Text Extractor To Work
Text analysis is no longer an exclusive, technobabble topic for software engineers with machine learning experience. It has become a powerful tool that helps businesses across every industry gain useful, actionable insights from their text data. Saving time, automating tasks and increasing productivity has never been easier, allowing businesses to offload cumbersome tasks and help their teams provide a better service for their customers.
If you would like to give text analysis a go, sign up to MonkeyLearn for free and begin training your very own text classifiers and extractors – no coding needed thanks to our user-friendly interface and integrations.
Reach out to our team if you have any doubts or questions about text analysis and machine learning, and we'll help you get started!