Time Travel: 2014

Moreover, the format of this academic exchange is very similar to Eve Carly’s previous team exchange method.

Several people with similar interests gather together and start chatting.

All this gave Eve Carly a rare sense of intimacy in a foreign country.

Eve Carly was flattered by these thoughtful arrangements.

Apart from these, what Eve Carly cares about most is Lin Hui's academic attitude.

What surprised Eve Carly the most was Lin Hui’s academic attitude.

Although Lin Hui is far ahead of others in terms of research results, Lin Hui has no academic pretensions at all.

Lin Hui is very good at listening.

This is an extremely rare thing.

It seems more difficult to find an expert who is good at listening among natural language processing experts and scholars than to find a giant panda on the earth.

Many natural language processing experts and scholars are computer practitioners.

In Eve Carly's impression, this kind of people have always expressed themselves, and what they are least good at is listening.

Maybe it’s not that I’m not good at listening, but I simply don’t like listening.

It seems that listening to other people's thinking lines and research status can easily remind them of the painful experience of debugging in their early years.

But the situation is very different here in Lin Hui. Lin Hui is very good at listening.

At the beginning of the symposium, Eve Carly originally planned to let Lin Hui express his opinion first.

As a result, Lin Hui signaled her to speak first, which made Eve Carly very uncomfortable.

For a while, she didn't know what to say, so she could only elaborate on the email she sent to Lin Hui not long ago.

The reason why Eve Carly introduced this aspect was because she couldn't think of a suitable starting point for the topic.

There is another reason, that is, Eve Carly is curious about how Lin Hui evaluates text similarity when building the LH text summary model.

But Eve Carly was too embarrassed to ask this question directly, so she had to make insinuations.

Eve Carly was a little nervous when she first started to explain, and was afraid that repeating the repeated content in the email would cause Lin Hui's dissatisfaction.

But Lin Hui didn't seem to mind and listened to her statement seriously.

Lin Hui's attitude made Eve Carly less nervous.

When making the statement, Eve Carly noticed a small detail:

On the way back from the airport, when she suggested to Lin Hui that she find an interpreter for simultaneous interpretation, Lin Hui agreed almost without thinking.

But during the actual communication, Eve Carly judged from some of Lin Hui's reactions that Lin Hui could actually directly understand what she expressed.

In this case, why did Lin Hui agree to her request in the first place?

Instead of just throwing away the translation and communicating with her?

Maybe all this is to give equal respect!

This equal respect is not only given to Eve Kali, but mainly to Mina Kali.

Imagine if Lin Hui could talk to Eve Carly without the need for an interpreter.

It seemed that the most embarrassing person was Mina Carly who was traveling with Eve Carly.

It is indeed not easy for a man to be so careful.

Eve Carly's favorable impression of Lin Hui increased by a few tenths of a percentage point.

My sister seemed to have noticed Lin Hui’s considerate approach in simultaneous interpretation.

Eve Kali noticed that Mina Kali had flipped her hair several times intentionally or unintentionally.

Of course, Mina's overture might just be because of Lin Hui's appearance.

It stands to reason that the appearance of Eastern men is difficult to distinguish in Western eyes.

But handsome people transcend geographical limitations to a certain extent.

This seems to be the case for Lin Hui. Even if judged by the most demanding aesthetic system, Lin Hui's appearance can be rated 99 points out of 10 points.

When she first saw Lin Hui, Eve Kali even thought that Lin Hui would be a model if Lin Hui had not taken the initiative to reveal his identity.

Of course these are digressions.

After noticing Lin Hui’s kindness in details, intentionally or unintentionally.

Eve Carley completely relaxed as she made her statement.

Xiang Linhui focused on how people evaluate text similarity in this time and space.

Eve Carly noticed that Lin Hui frowned when she heard that her team had previously used a method based on network knowledge to evaluate text similarity.

Could it be that Lin Hui doesn't agree with the method of evaluating text similarity based on network knowledge?

Or does Lin Hui think there is any better way than this method?

Eve Carly kept this in mind silently.

After Eve Carley's statement was completed.

Lin Hui understood what she meant.

However, he did not answer Eve Carly's question directly.

Instead, I asked Eve Carly: "What do you think about using vector intervention for semantic text similarity calculation?"

Although this was the first question Lin Hui raised in this exchange.

But this question caught Eve Carley somewhat off guard.

Eve Carly wasn't sure why Lin Hui raised this question.

Is it possible to calculate semantic text similarity without relying on vectors?

But how can this be done?

When the machine recognizes text, in order for the machine to recognize natural language, the natural language is often digitized.

To distinguish these values ​​​​by attributes, vectorization must be performed.

This method has a long history. Eve Carley remembers that researchers first proposed the vector space model VSM in 1977 (this time and space).

This research method became more popular once it was proposed.

Although it was soon discovered that this method had quite a few loopholes.

Using the VSM method, when the amount of text is large, the generated text vectors are very sparse, which leads to a waste of space and computing resources;

In addition, VSM ignores the relationship between words in order to achieve the effect of simplifying the model. In many cases, there is a connection between words, so it is unreasonable to simply think that words are independent of each other.

Despite the obvious loopholes, in the following nearly four decades of history, people still introduced vectors for semantic text similarity analysis.

Take Eve Carley's previous team, although they previously used a method of calculating text similarity based on network knowledge.

But essentially it just maps the web content in Wikipedia into high-dimensional vectors.

Then the semantic text similarity is calculated using a vector space-based method.

It can be said that it still cannot leave the shell of vector space.

Although forty years later, the so-called "waste of space and computing resources" encountered back then can be solved to some extent through hard stack computing power.

But this can only solve the problems encountered back then.

The amount of information and complexity faced when processing text now is completely different from that of the past.

At this time, vectorization is facing a new difficulty - dimensionality explosion!