When the computer understands your emotions, will the world get better?

When the computer understands your emotions, will the world get better?

Some companies and researchers are trying to use computers to understand the emotions behind the text: Although sentiment analysis products are far from perfect, they have been able to extract something from big data, and even one day in the future can monitor human mental health .

Many people think that 2020 is the worst year in history. Although this description may be too subjective, there is a data to support this conclusion.

Hedonometer (Translator’s Note: Hedono is a root, meaning pleasure) is a computerized evaluation method that detects our happiness and loss. It runs day after day on a computer at the University of Vermont, collects about 50 million tweets from Twitter every day, and then performs a quick and rough interpretation of public sentiment. According to Hedonometer data, 2020 is its worst year since it started recording in 2008.

When the computer understands your emotions, will the world get better?

Sentiment analysis has been used in a variety of scenarios. Image source: pexels

For more than 50 years, computer scientists have been studying how to use computers to evaluate the emotional tone of words. Hedonometer is a relatively new development they have made. In order to build Hedonometer, Chris Danfoss, a computer scientist at the University of Vermont, needs to teach machines to understand the emotions behind these tweets. After all, it is impossible for humans to interpret all tweets one by one. This process is called sentiment analysis, and it has made significant progress in recent years and has a variety of usage scenarios.

In addition to perceiving the emotional state of Twitter users, researchers also use sentiment analysis to study people’s perceptions of climate change and to verify common sense, such as whether minor chords are more sad (and the degree of sadness) than major chords in music , and so on. Some companies that covet emotional information from customers are using sentiment analysis to evaluate reviews on platforms such as Yelp (the largest review site in the United States) , and some companies are using it to perceive employees’ emotions on internal social networks at work. This technology may also find applications in medicine, such as identifying patients with depression who need help.

Danfoss said that sentiment analysis can help researchers analyze a large amount of data, which was difficult to collect in the past and the process is also time-consuming: ” In social sciences, we tend to measure easy things, such as gross domestic product. And happiness itself is a very important but difficult thing to measure .”

1. How to read your words

You might think that the first step in sentiment analysis is to teach computers to understand what humans are saying, but this is something computer scientists cannot do. Understanding language is one of the most notorious problems in artificial intelligence. In fact, there are a lot of emotional clues behind the written text. Even if you don’t understand the meaning of the text, the computer can recognize the emotion.

The earliest sentiment analysis method is word counting. The idea is very simple. It is to count the number of positive words and the number of negative words. A better approach is carried out according to the meaning of the word weighting, such as “excellent (Excellent) ” than “good (Good) ” expressed stronger feelings, these weights are usually experts configurations: a portion of sentiment analysis is often used corresponding to the emotional word dictionary, this method is called lexicon method (lexicons) .

When the computer understands your emotions, will the world get better?

The simplest method of sentiment analysis is the thesaurus method. Image source: pixabay

However, only counting the number of words has its inherent problems. One is that the word order is ignored, and the sentence is only regarded as a collection of words. In addition, the lexical method may miss some clues from specific contexts, such as this product review: “I’m so happy that my iPhone is nothing like my old ugly Droid. (I’m very happy, my iPhone and my old ugly Droid Android phones are completely different.) “This sentence has three negative words (“completely different”, “old”, “ugly”) , and only one positive word (“happy”) ; although humans can immediately realize ” “Old” and “ugly” refer to different mobile phones, but for computers, these are both negative. At the same time, the comparative context will bring more difficulties. What does “completely different” mean? Does the speaker want to compare the two? Language can be confusing at times.

In order to solve these problems, computer scientists have increasingly turned to more complex methods to completely exclude human labor from this process. They are using machine learning teaching applications to recognize some paradigms, such as meaning relationships between words. For example, the computer can learn that when the two words “bank” and “river” often appear together, “bank” means “river”, and when “bank” and “money” appear in the same sentence, then It might mean “bank”.

2013 machine learning achieved in this regard significant progress, Google Brain researcher Thomas – Miko Rove built called word embedded (word embeddings) tool, this tool will map each word to 50-300 numbers A list, called a vector. These numbers are like fingerprints that describe a word, which can describe its characteristics when it appears together with other words.

In order to obtain these descriptors, Mikrov’s program collated millions of words in newspaper articles and tried to predict the next word given the previous word. Mikrov’s embedding method can recognize synonyms: words like “money” and “cash” have very similar mappings. What’s more ingenious is that although this tool can’t actually recognize the meaning of these words, it can capture some basic analogies: for example, the king is to the queen as the boy is to the girl. Given that this analogy is the scope of the American College Entrance Examination (SAT) , it is already a remarkable achievement to be able to do so.

Mick words embedded Rove is a hidden layer having a (Translator’s Note: The input feature data into a certain neuronal structures dimensional space to another, so that it can be reasonably classified by) neural network generated. In recent years, the neural network using the human brain as a loose model has made machine learning amazing progress, and its outstanding representative is AlphaGo. Mikrov’s network is a specially designed shallow network that can be used in a variety of scenarios such as translation and topic analysis.

A deeper neural network has more “cortex” that can extract the emotional information of individual words in the context of a specific sentence or document. A common reference task is to have a computer read film reviews on the Internet Movie Database (IMDb) and predict whether the reviewer will give a good review or a bad review. The lexicon method first reached an accuracy rate of about 74%, and later more complex methods only reached an accuracy rate of 87%; the earliest neural network method achieved a score of 89% in 2011, and their accuracy rate is now It is as high as 94%, which is close to the human level . However, humor and irony are still big stumbling blocks, and language expression may be contrary to expected emotions at this time.

Despite the many benefits of neural networks, the thesaurus-based approach is still very popular, and Danfoss has no intention of changing his Hedonometer thesaurus. Neural networks may be more accurate in terms of the results of some problems, but they also come at a price. The training period of machine learning alone is already one of the most complex tasks a computer can run.

” Basically, it’s limited by how much electricity you have ,” said Robert Steen of the Wharton School of Business, who introduced the evolution of sentiment analysis in the “Annual Review of Statistics and Its Applications in 2019.” “How much electricity does Google use to train AlphaGo? The joke I heard is that these electricity are enough to boil the ocean.”

In addition to power requirements, neural networks also require expensive hardware and certain expertise, and the process of machine learning lacks transparency. The computer is figuring out how to handle tasks on its own, rather than following the programmer’s instructions step by step. As a pioneer in the field of sentiment analysis, Professor Liu Bing from the University of Illinois at Chicago also said that it is easier to correct mistakes using thesaurus method.

2. Measuring mental health

Although sentiment analysis is usually within the purview of computer scientists, it has deep roots in psychology. In 1962, Philip Stone, a psychologist at Harvard University, developed General Inquirer, which was the first computer general-purpose text analysis program for psychology. In the 1990s, social psychologist James Pembeck developed an early sentiment analysis program using linguistic surveys and word counting to observe people’s psychological world.

These early assessments revealed and confirmed the long-term observations of experts that people with depression have a unique writing style : for example, they will use the pronouns “I” and “me” more often, and use more negative words and speech. There will be more words related to death.

When the computer understands your emotions, will the world get better?

Patients with depression have a unique writing style. Image source: Pexels

By analyzing social media posts, researchers are exploring the state of mental health expressed in speech and writing. Danfoss and Harvard University psychologist Andrew Reese analyzed the Twitter posts of patients with depression or post-traumatic stress disorder before being diagnosed, and he found that signs of depression began to appear as early as nine months ago. Facebook has a special algorithm to detect users who are at risk of suicide, and human experts will review these cases, and if necessary, will send reminders to users or provide a hotline number.

However, social network data is still a long way from being used for patient care. Privacy is an obvious issue, and more work is needed to prove its effectiveness. Steve Chansler, an expert on human-based computing (translator’s note: a system engineering methodology that combines computer science and social science) at Northwestern University , is the co-author of a review report of 75 such researches. Many studies that assess mental health fail to correctly define their terms, or do not provide enough information to repeat the results. But she still believes that sentiment analysis is helpful for medical treatment, such as triage new patients. And even without personal data, sentiment analysis can identify trends, such as the overall stress level of college students during the COVID-19 period, or the types of social media interactions that trigger repetitive eating irregularities.

Three, read emotions

Sentiment analysis is also used to solve some easy problems. In 2016, Nick Obradovic of the Max Planck Institute for Human Development in Berlin analyzed about 2 billion tweets on Facebook and 1 billion tweets on Twitter. , Studied the influence of weather on mood. 25mm of rain reduced people’s happiness by about 1%, while sub-zero temperatures reduced their happiness by about 2% . In a follow-up study, Obradovic and his colleagues used Twitter to understand how people feel about climate change and found some depressing results. They found that after five consecutive years of global warming, the general awareness of Twitter users has changed: they no longer talk about climate warming on Twitter. Nevertheless, the data shows that users’ happiness is still affected by it. “It’s like boiling a frog in warm water,” Obradovic said. “This is one of the most disturbing empirical findings of all the papers I have done.”

Monday is notorious for being the worst day of the week. Nevertheless, Danfoss’ Hedonometer’s early analysis of tweets found that Tuesday was actually the day when people’s emotions were at their lowest. Of course, Friday and Saturday are the happiest days, so naturally there is no need to talk about it. However, after the 2016 U.S. election, the mood pattern has changed every week. Although the weekly emotional cycle is still retained, people’s attention is attracted by other things beyond this, and the topic level even exceeds the general elements of life. Danfoss said: “On Twitter, political topics never stop. Any day of the week can be the most frustrating day.”

Another theory that has been tested is that in music, major chords are considered more cheerful than minor chords . An Yongye, a computational social science expert at Indiana University, tested this theory by analyzing the emotions of the lyrics accompanying each chord in 123,000 songs, and found that major chords are indeed closely related to happy lyrics, with a score of 6.3 on a 9-point scale. , Surpasses the minor chord’s 6.2 points. Although this difference seems insignificant, in Hedonometer’s measurement dimension, the emotional difference between Christmas and normal working days is only 0.2 points. An Yongye also compared different music genres and found that the rock music of the 1960s was the happiest emotionally, while heavy metal was the most negative.

Four, business acumen

Sentiment analysis is also widely used in the business world, but many companies do not discuss it publicly, so it is difficult to accurately measure the extent of its audience. Liu Bing introduced: “Microsoft, Google, Amazon…every company is doing it, and some companies have multiple research groups.” An easily available measurement standard is currently publicly available commercial and academic sentiment analysis programs. quantity. A benchmark comparison analysis in 2018 detailed 28 such procedures.

Some companies use sentiment analysis to understand what their users are saying on social media. There is a seemingly fanciful example: Canadian airline Expedia Canada launched a marketing campaign in 2013, but the response was counterproductive. People hate the harsh violin background music in the advertisement. Expedia soon released a new video mocking the old ad. They invited a dissatisfied Twitter user to smash the violin. People think that Expedia gets social media response through sentiment analysis. Although this is difficult to prove, it is certainly something that sentiment analysis can do.

Some other companies use sentiment analysis to track employee satisfaction by monitoring their internal social networks. IBM developed a program called Social Pulse to monitor the company’s internal network to understand what employees are complaining about. Due to privacy reasons, the software only viewed posts that were public throughout the company. Even so, this trend still troubles Danfoss: “I am worried that the company’s bottom line exceeds the privacy of employees, which is not beyond blame in terms of ethics.”

As sentiment analysis becomes more common, moral issues may become the next hidden worry. Companies, mental health experts, and any other subjects considering the use of sentiment analysis should keep in mind that even if sentiment analysis has a promising future, realizing this ideal is still fraught with risks. The mathematical knowledge that the analysis relies on is the relatively easy part, but the hard part is actually understanding humans. As Liu Bing said, “We don’t even understand what true understanding is.”

This article is authorized to be translated from Knowable Magazine, a magazine under Annual Reviews, with the original title ” How algorithms discern our mood from what we write online “, author Dana Mackenzie, published on 2020.09.14.

Posted by:CoinYuppie,Reprinted with attribution to:https://coinyuppie.com/when-the-computer-understands-your-emotions-will-the-world-get-better/
Coinyuppie is an open information publishing platform, all information provided is not related to the views and positions of coinyuppie, and does not constitute any investment and financial advice. Users are expected to carefully screen and prevent risks.

Like (0)
Donate Buy me a coffee Buy me a coffee
Previous 2021-07-20 08:22
Next 2021-07-20 08:46

Related articles