@Scientific research party, this is probably the most useful paper reading tool, and it's free
Yuyang comes from Aofei Temple
Qubit | Official account QbitAI
My fellow Taoist, I wonder if you also have the trouble of being inefficient in reading English papers while you are reading salt?
As an English scumbag seen every day on arXiv, I can’t extricate myself from the road of discovering artifacts for paper reading.
Not recently, another desktop translation software has been planted by netizens.
Let's try it out. The result of translating PDF is like this:
There is also a sentence-by-sentence comparison function:
If you just want to take a look at the summary, you can just take a screenshot. There are also Chinese and English translations:
My mom no longer has to worry about me deleting so many times that my hands get cramps (dog head).
△ The text copied directly from PDF has extra line breaks, which affects the translation effect
This function comes from the recently updated NetEase Youdao Dictionary 9. To be honest, the effect is really good at first glance~
If you further browse the official website of Youdao Dictionary, you will find this evolution. This tool emphasizes improving " academic productivity ."
As to whether this is really the case, let's try it carefully together.
The entire process from paper retrieval to paper writing is available
Reading English documents and translating PDFs is a necessity.
Previously, the academic helper Google Translate had a PDF document translation function.
However, Google's document translation always has some minor issues.
For example, the layout is prone to confusion, which makes it seem like you can understand every word, but you don’t know what it means when put together.
For example, the original table looks like this:
Use Google Translate and you will get a result like this, where the words are misplaced and somewhat difficult to understand:
In comparison, the results given by Youdao look like this:
It can be said that Youdao Dictionary is obviously better in terms of layout maintenance.
What’s even more convenient than Google Translate is that Youdao Dictionary also has an intuitive bilingual comparison function.
After all, although machine translation can improve reading efficiency, it is still unrealistic to read the paper completely without referring to the original English text.
Compared to the hectic operation of transferring two documents back and forth, it is much more convenient to compare them directly on the same screen.
What’s even more interesting is that Youdao Dictionary’s “academic evolution” has made targeted optimizations to professional fields such as computer science, medicine, and financial economics , which can improve the accuracy of term translation.
Custom glossaries are also supported .
For example, if you find that the word "self-attention" is not translated in the article, you can directly select this word and choose Term Correction to add it to the term base. After applying the term base, Youdao Dictionary will keep the same translation when encountering the same term again.
In addition to making the translation work more convenient and accurate, unlike other translation tools, Youdao Dictionary 9 even plans to take care of the entire process of writing papers for scientific research parties this time .
They work directly with ReadPaper and Aminer, with built-in paper retrieval services. The official statement is:
Covering more than 40 professional fields and containing 300 million documents.
△ Search for highlights
During the search process, you can also turn on the bilingual comparison button to quickly read the summary.
That’s not all. Youdao Dictionary 9 can also come in handy when you are writing a paper: its “AI English Writing Correction” function can intelligently identify grammatical errors, provide example sentences for reference, “native-level” sentence polishing, and automatically mark the source of citations and generate reference document information…
The technical secrets behind the "academic" dictionary
Youdao Dictionary 9's one-stop academic service is truly aimed squarely at researchers.
Behind these functions that are closely aligned with academic needs, the technical challenges are also worth talking about.
PDF document translation: PDF parsing algorithm based on visual technology
Let’s talk about the most eye-catching PDF document translation function first.
PDF document translation itself is a relatively complex problem, involving multiple processing links, including PDF document parsing, Word document rendering, translation strategy, translation result backfilling, etc.
Especially for unparsable PDF (scanned PDF), the text, vectors, pictures and other elements in the page are actually included in the same picture and cannot be directly parsed.
At this time, you need to use OCR technology to identify the various basic elements and then convert them into editable documents.
The difficulty here is that compared with simple text recognition, OCR recognition of paper documents is a more difficult task. Because it is not just text, but also tables, images, and formulas in the paper. If they cannot be accurately recognized, it will easily lead to content loss and layout errors.
In addition, in order to keep the layout consistent with the original text, AI needs to accurately determine the single column, multi-column, header, footer and other information of the page, so that it is possible to restore the original layout and content in the editable document.
For unparseable PDFs, Youdao has made improvements in two aspects.
In terms of text recognition , Youdao's self-developed OCR algorithm integrates the industry's leading text detection and recognition models, which can achieve a higher text recall rate and cover dozens of major languages.
In terms of layout recognition , a "divide and conquer" strategy is adopted.
Specifically, after obtaining the basic layout information through the layout analysis model, AI will adopt different extraction strategies based on different composition results.
For example, when encountering a common layout that is mainly text-based and rich in tables, Youdao Dictionary will mainly use text segmentation algorithms and table analysis algorithms to process the document.
For special layouts that are mainly based on images and have complex backgrounds, all elements of the original layout are retained as they are, focusing on text processing, and finally performing multi-layer overlay rendering.
In addition, Youdao has also optimized many details to make the final layout presented to users more accurate and regular.
For example, the transparency processing of text boxes and the insertion rules of line breaks.
Professional field translation model
As mentioned earlier, in addition to general translation, in Youdao Dictionary 9, users can also choose professional fields such as medicine and computers to further improve the accuracy of translation.
Behind each professional field, there is actually an independent translation model.
Specifically, the technical team collected a large amount of data from different professional fields and conducted customized training and enhancement of the neural network translation model.
In this way, even terms like Pneumonoultramicroscopicsilicovolcanoconiosis can be accurately translated.
In addition, generally online machine translation algorithms are based on sentence units. To meet the needs of document translation, Youdao Dictionary 9 has introduced a chapter algorithm this time . In other words, when AI performs translation, it will combine contextual sentences to improve the translation quality.
From this point of view, NetEase Youdao Dictionary 9’s claim to be the “most powerful desktop translation software” is not just bragging, but rather well-prepared.
In fact, in addition to new attempts in technical details, there is also a longer-term technological accumulation behind the transformation of Youdao Dictionary into a "scientific research artifact".
For example, in terms of neural network translation technology, NetEase Youdao started to develop its own neural network translation model YNMT in 2016, and is one of the earliest teams in China to develop its own machine translation technology.
In terms of OCR technology, NetEase Youdao launched the OCR engine of convolutional neural network + recurrent neural network in 2017. It can now support the recognition of more than 100 languages, and also has the ability of automatic language discrimination and mixed recognition.
Of course, the most important thing for us is that it is free and easy to use , and you won’t lose money if you invest in it.
Starting from September 13, the free version of document translation is available to all users.
Finally, if you have any other treasures of scientific research artifacts, please share them in the comment section~
-over-
click here