Presentation of a study comparing linguistic differences between native and non-native English speakers in 50 quantum-physics papers. The presenters extracted text from abstracts, introductions and conclusions, cleaned and tokenized it, analyzed lexical/part-of-speech features (nouns, verbs, adjectives, adverbs, pronouns, prepositions), and trained a classifier that achieved 84% accuracy distinguishing native vs non-native writing. Key findings: natives use more adjectives and adverbs and longer sentences; non-natives use more pronouns and (slightly) more prepositions; classifier performed better predicting non-native phrases.
Chapters
Welcome.
Our topic is comparison of papers of English and non English speakers in physics. We are Jovan Costiuk and.
So we studied the linguistic differences between native and non native English speakers in 50 physics papers. And we developed a classification algorithm obtaining an accuracy of 84%. The data comes from articles about quantum physics. And the main difficulty here was to get the data and classify that as a data data coming from a native and a non native. So the main question is what's the difference between a text written by a native speaker and one written by a non native speaker? Our goal is to understand how people share their ideas, how deep the understanding of a new language must be to communicate an idea and whether a text might have enough differences to be classified as a text written by a native.
Speaker or a non native speaker.
To explain in a general way what the article is about, we want to know the differences between the structures used when a native speaker and a non native speaker writes a scientific paper who uses more complex structures and how unique are the words for for that we got the text from the abstract introduction and conclusion. Why? Because they are less technical and use more everyday language.
So this is for our. This is our working directory. So can Just as you see, we.
Divided all the articles in a native path and another native path with a total of 25 articles of native and non native people. The data comes from articles about quantum physics. We highlighted the abstract introduction and a conclusion. And we created two functions to import.
And visualize the function.
The first one, get data that iterates over each path, importing only the highlighted article text, which is the text we want to analyze and show that shows a snippet of the element text.
So here is the native data. And here we can see it's the non native data. Here in the cleaning the data, what.
Do we need to remove? We need to check the structure of the phrases and the function fontling removes the parenthesis special symbols and digits.
So first we lower lowercase R our data. So as we found some math expressions, we need to translate math expressions into ASCII code. For example, an alpha will turn into alpha. So we here we translate into ASCII code. Then we delete the parentheses, then the.
Special characters that are not in the.
Format UTF8 and then we delete the numbers.
So after cleaning the data.
We got this.
This is our final result. So what is the next thing to do? Well, of course in our process was to combine all the text into a single string.
Here we can see it and we go. We defined the function get words.
To get the lexical features and count the data.
Here we get the lexical features and with this function length we count the.
Data and get the nouns, verbs, adjectives, etc.
So here we got all the data. Nouns, verbs, objectives, adverbs, pronouns.
What's the hypothesis? Our hypothesis is we can discriminate between text written by native and non native speakers, English speakers in the field of physics, by analyzing the lexical features. This is our hypothesis. So, starting with the graphs. So first.
This is our comparison of.
The parts of speech used in articles.
And their frequency by native and non native speakers. So we can see in this image, let's make it bigger. Here the native natives are in yellow and non natives are in blue.
In the xx axis is the structure.
And on Y axis is the number of words.
So we can see that natives use.
More adjectives and more adverbs than non natives and non natives use more pronouns and as likely more prepositions, but just a thin difference. Our second graph is the average composition of an article without duplicated information. So let's make it bigger here. So the average composition on a native that are inside and non native that are outside. So we can see that adverbs, prepositions.
And punctuation is almost the same, but.
Natives use fewer nouns that are in purple, verbs and adjectives. So that's the main difference here.
What's the percentage of.
Repeated words on each category?
So.
We define it the function rep. To get the percentage and let's make it bigger. Native are in purple, non native are in blue. So we can see that non natives use more verbs, adjectives and natives use more adverbs.
So these are. This is according to our last.
Slide.
This is according to this slide because.
We had the same data as the past slide.
What's the percentage of structure used in the text?
So we defined the function rat that takes two arguments, the kind of lexical feature and the total.
Phrases in the bear in the text.
Sorry. So non natives are outside and natives are inside. So again, punctuation, prepositions and adverbs are almost the same. But we have a difference here between verbs that are more used by natives and adjectives.
What's the length of the sentence?
So we divide orglata and got the length of the sentence with this instruction. So and then we did an histogram. So where native are in yellow and non native are in blue. So we can see that for example here the length of the sentence created by native people are larger than sentences created by by non native people.
So in conclusion, sentence sentences made by natives could be harder to understand.
Why? Because they are larger, they have more contents.
Now the question will be, now the question is, will it be possible to.
Tell whether a sentence belongs to a native speaker or a non native speaker.
Just by looking at the words using in the sentence? We split our data and created a classifier. So first we split the data into.
Training and test set. Here we split and we divided that into training and test set and then we use the function classify and it automatically, automatically identified the input as text. So we have this the the classifier and we can see that native classified as native was 15 examples, native classified as non native 2 examples, non native classified as native were 3 examples and non native classified as non native were 11 examples.
So we got an accuracy of 84%.
So checking the false negative rate and false positive rate and the precision, we can see that non native speakers were.
More.
Precise than native speaker native speakers. So our conclusions are the way that we express our ideas is through the words we use. With this word we could identify many points. For example, the use of adjectives and adverbs is one of the main differences between native and non native speakers. More words are used in a sentence with a native speaker. And that's why I think it's harder to understand these kind of phrases.
The most common parts of speech are adjectives and nouns. The expressions used in the articles were different enough to get an 84% chance.
Of classifying whether a sentence was written by a native or a non native English speaker.
And it was easier for the classifier to predict a phrase, to predict if.
A phrase is from a non native speaker.
So that's all.
Thank you.
Hello people, do you have any question you can ask? Thanks, Thomas. Yes, it was fun. We had a lot of fun making this and getting the data and talking about the topic, how we decided the topic. And it was really, really nice. Thanks for your comment.
