Time Travel: 2014

LIN HUI's algorithm is a crushing leader in terms of actual performance efficiency and algorithm theory, not just a little bit ahead.

If the lead is only a little bit, it may be just a fluke, but a crushing lead means an unquestionable gap in strength.

Eve Carly has a deep understanding of this. The generative summary algorithm proposed by LIN HUI has even pioneered some unprecedented research in subdivided fields.

A new segment that is easier to know is sentiment analysis of natural language processing.

This direction is a completely new topic for Eve Kali, who has only conducted research on extractive text summarization before.

But this is far from all. The generative summary algorithm proposed by LIN HUI should also involve many groundbreaking things.

However, due to the inability to see clearly, Eve Carly is still unclear about how much groundbreaking research is involved in LIN HUI's algorithm.

Although it is unclear how many new segments are involved.

However, Eve Kali relied on the intuition of a scientific researcher and conservatively estimated that LIN HUI's algorithm would involve five or six groundbreaking research in subdivided fields.

However, it was precisely because of this that Eve Carly couldn't understand it.

Obviously LIN HUI's algorithm is so far ahead, why is it still seeking to acquire many "low-level" patents?

Eve Carly looked at the patents on the patent website that LIN HUI Gray was seeking to acquire.

They all seem "low-level".

When it comes to the distinction between low-level and high-level, the patent itself does not distinguish between low-level and high-level.

However, in a research system, there will be low-level and high-level research due to different divisions of responsibility and different levels.

After LIN HUI proposed a generative text summarization algorithm.

Current automatic summarization implementation methods are mainly divided into extractive methods and generative methods.

After recent research on the technical route of generative summarization algorithms, Eve Carly has been able to easily understand the difference between generative text summarization and traditional extractive text summarization.

The so-called extractive summary extracts key text units from the original document to form a summary.

Generative summary forms a summary based on the understanding of the input original text. The generative summary model tries to understand the content of the text and can generate words that are not in the original text, which is closer to the essence of the summary and has the potential to generate high-quality summaries.

Although there are slight differences in the specific summarization between the two, both are essentially automatic text summarization.

Since both are automatic text summarization, the technical frameworks of both can be summarized as:

Content representation → weight calculation → content selection → content organization.

ⅠContent representation is the process of dividing the original text into text units, mainly preprocessing work such as word segmentation, words, sentences, etc.;

The main purpose of content representation is to process raw text into a form that is easy for algorithms to analyze through preprocessing.

II Weight calculation is to calculate the corresponding weight score for the text unit (i.e., the original text after preprocessing). There are various ways to calculate the weight, such as calculating the weight based on feature scores, sequence annotations, classification models, etc. to extract content features.

The purpose of this step is to complete a preliminary analysis of the preprocessed original text through this series of calculations.

III content selection is to select the corresponding subset of text units from the weighted text units (that is, the text analyzed by the weight in step II) to form a summary candidate set, which can be based on the required summary length, linear programming, submodular function, and heuristics. Formula algorithm, etc. to select text units;

IV Content organization refers to organizing the content of the candidate set to form a final summary, which can be output in order according to the word count requirements. Some researchers have also proposed using methods based on semantic information, templates and neural network learning to generate summaries that meet the requirements.

(ps:...In popular terms, II is a bit similar to finding the key paragraphs when summarizing an article reading;

III is similar to the process of further finding key sentences and keywords based on II;

IV is similar to the process of determining key sentences and keywords and then using appropriate language to form a final reading summary)

Judging from the corresponding descriptions of these levels of the technical framework, we can see that weight calculation, content selection, and content organization are all very important.

If you can't figure out the weight calculation and content selection, you won't know exactly where to summarize the text when summarizing.

After all, not everything in an article is the focus. Just like when we summarize the reading, we usually focus on the first and last paragraphs and the beginning and end of each paragraph. It can be said that the weight calculation and content selection play a role. It is to determine where the corresponding text feature points of the text to be processed are mainly concentrated.

If you can't organize the content, you won't be able to get a reasonable and smooth text summary even if you can find the densest text features.

Content representation is less important than these three.

Because of the above-mentioned division of labor, if things related to text summaries need to be hierarchical.

Then the patents related to weight calculation, content selection, and content organization can be said to be advanced patents in the text summary system.

Patents related to content representation are low-level patents.

"A New Method for Text Judgment, Screening and Comparison" is essentially used to screen original texts.

According to the introduction just now, this undoubtedly belongs to the level of content expression.

Just such a "low-level" patent,

Eve Carly really doesn’t understand why LIN HUI is seeking to purchase such a patent.

Could it be that as Ewald Cherry said, what LIN HUI cares about is not the "A New Method for Text Judgment, Screening and Comparison" algorithm patent itself.

What LIN HUI cares about is her who got the patent?

How can it be? If you are just interested in her, you can contact her directly.

After all, it is easy to find the contact information of scientific researchers.

Eve Carly guessed several possibilities, but could not guess a reasonable explanation.

…

Although it's a bit hard to understand.

But in the end, Eve Carly sold the patent "A New Method for Text Judgment, Screening and Comparison" to LIN HUI.

After all, the only question the seller has to think about is whether the buyer’s offer is appropriate, not why the buyer is buying.

In addition, Eve Carly noticed that one of the patents Lin Hui had previously purchased had been successfully transferred for US$500,000.

Although the value of this patent was not as high as hers, the seller was Asile Velasquez.

Eve Carley remembers this man as a senior researcher at Google Research (Google Search, the predecessor of Google AI).

To be honest, the reason why Eve Carly was struggling with whether to sell this patent was not because she was worried about making less money, but mainly because she was worried about the negative impact.

But now no one inside Google is worried about the negative impact caused by the patent transfer, so what else does she care about.

As for the question of whether the patent "A New Method for Text Judgment, Screening and Comparison" has potential value, it is no longer important.

Eve Carly is quite looking forward to this patent, which she has never seen any value in, taking on new life in the hands of LIN HUI.

Perhaps LIN HUI will prove that there are no low-level patents, only low-level vision.

Thanks for reading (●ˇˇ●)