Chapter 148: You Think You Understand Machine Translation Better Than Me?
“Professor, you’re not doing language translation. Language is a game of rules; probability is too unreliable.” Paul Calvin still wanted to struggle a bit more.
Of course, he truly didn’t believe that translation had any relationship with statistics either.
Word-for-word correspondence.
English words and Russian words correspond one-to-one, direct literal translation, expand the corpus.
In the thinking of that time, this was the right path.
That is, the so-called exhaustive method.
After making one-to-one correspondence for all words, automatic translation would be achieved.
Statistics, probability games—not to mention if Lin Ran was right, their incompetence would be exposed completely; just the improvement principle mentioned by Lin Ran was intuitively wrong.
Simply put, counterintuitive.
Just like before the GPT large model came out, everyone thought algorithms were the most important.
After GPT came out, everyone rushed to brute-force with massive computing power.
By the time of DeepSeek, it seemed algorithms were somewhat useful.
Even top researchers would have issues with blind conformity, confusion, inability to find direction, and being stuck.
In this chaotic era of computers, it was perfectly normal.
“Precision? Precision means errors; current computers are far from reaching precision.
Don’t you know that the good effect you demonstrated in ’54 was because those Russian sentences were carefully selected by you?
The complexity of actual natural language far exceeds your expectations.
You’ve only done corpus expansion; you haven’t done rule coverage or context dependency processing.
Do you understand machine translation better than I do?”
Lin Ran roared: “You’ve made no progress in nine years; now immediately do as I say!”
Lin Ran’s status, strength, and power were right there; they had no choice to refuse.
Whether Watson believed in Lin Ran aside—after all, the Deep Blue project had just ended—Defense Department McNamara followed whatever Lin Ran said.
Could you computer guys understand computers better than a mathematics master?
McNamara hadn’t forgotten the flair Lin Ran showed in game theory and statistics.
IBM’s CEO supported Lin Ran, the Defense Department minister supported Lin Ran; the Georgetown University research team could only be rubbed on the ground.
“We have five points in total: optimization algorithm and rules design, corpus and vocabulary expansion, improved data processing efficiency, introduction of statistical methods and maximum hardware utilization.
Among them, improving data processing efficiency and maximum hardware utilization are handled by the IBM side.
The other three points are handled by the Georgetown University members.
Let’s first talk about optimization algorithm and rules design.
Your persistent problem is that you haven’t introduced more refined syntactic rules in the expansion of the ruleset.
Because storage is limited, you thought expanding the bilingual vocabulary was enough.
In fact, syntactic rules are even more important.
You need to introduce common high-frequency patterns.
Perform dependency processing on context. Make vocabulary translation consider preceding and following words, reducing ambiguity through a limited context window.
For example, svet means both light and world.
This can be completely determined by the preceding word whether it’s light or world.”
Watson weakly reminded: “Professor, you know Russian too?”
Lin Ran looked completely natural: “Of course; I’ve met Korolev twice—how could I communicate with him without knowing Russian?
I know Russian language, German language, English language, and Chinese language.”
The identity of a multilingual master added credibility to Lin Ran’s theory.
In this era, scientists knowing several languages wasn’t strange.
Of course, some sensitive departments would increase suspicion of you.
Take John McCarthy mentioned earlier as an example: he was fluent in Russian, raised with Russian education from childhood, though born in America.
“Additionally, the translation process should be modular design, not simple mapping relationships.
It should be divided into preprocessing, translation, and postprocessing three parts.
Preprocessing includes tokenization and lemmatization; translation is dictionary mapping; postprocessing adjusts word order.
This reduces single computation complexity and improves rules reuse rate!”
Lin Ran’s words gave the research team members present a lot of inspiration.
It was like being trapped in the jungle of Vietnam without a way out before, and now a light appeared in the sky guiding them on how to escape the jungle maze.
Everyone was a bit eager to try it.
All researchers frantically recorded what Lin Ran said in their notebooks.
Though uncertain if the professor’s method would work, having a path was better than none before.
Moreover, if you didn’t record it well, being fired would just be a word from the professor.
“Alright, we just covered some simple content; now comes the hardest part.
Because IBM’s machines aren’t that powerful, we can only introduce some relatively simple statistical methods to improve our translation accuracy.
I call it frequency-based word alignment.
This is also the core of introducing our statistical model.
First, we manually analyze parallel sentences, annotating correspondence between Russian words or phrases and English translations.
Russian sentence: My govorim o mire
English translation: “We speak about peace”
Alignment result: “my” corresponds to “we”
“govorim” corresponds to “speak”
“o” corresponds to “about”
“mire” corresponds to “peace”
Then we need to count the frequency of such alignments.
Count the frequency of each Russian word or phrase’s corresponding English translation.
For example, in the corpus, “govorim” is translated as “speak” in 80% of sentences, 20% as “talk”.
This way, we can build a probability table.
Organize these probabilities into a table for the machine to query. Due to limited memory, we temporarily store only high-frequency word pairs, like the top 1000 by occurrence, ignoring low-frequency cases.
When translating a word with multiple choices, refer to the probability table to select the most likely translation.
Additionally, count co-occurrence frequency of adjacent words. My often appears with govorim, corresponding to We speak; the machine prioritizes this combination during translation.
Use rules for priority handling and statistical methods for ambiguous cases to compensate for rules’ shortcomings!”
Lin Ran gave them a good lesson from the perspective of statistics.
But this was just a beginning.
The research teams present learned the outline of Lin Ran’s optimization strategy; there were still many details to adjust, try, and optimize in the specific practice.
But just introducing probability as mentioned now gave the senior researchers of the Georgetown Translator a sense of sudden realization.
The earlier optimization algorithm and rules design sounded reasonable, but they couldn’t judge if it would really work in practice.
But introducing statistical methods—just imagining it—they knew it could significantly improve the Georgetown Translator’s effect.
After the day’s work ended, in a small diner near Redstone Arsenal, Calvin and Dostert sat in the corner with two glasses of local specialty beer in front of them.
Calvin put down his notebook, sighed, and said: “Leon, are we really idiots?”
After hearing today’s lecture, Calvin was doubting his life.
Lin Ran proposed a complete set of solutions; even if complete, many points they had thought of but didn’t know how to implement, plus some they hadn’t even thought of.
Nearly ten years of a whole team’s R&D ideas were no match for Lin Ran’s afternoon of practical insights.
Calvin was already doubting his life.
“The professor’s ideas aren’t ahead of time but too practical.
You feel they’re fanciful, but thinking them combined, they seem incredibly practical.
Even before starting, just from the framework proposed by the professor, I can imagine how good the effect will be after upgrading the Georgetown Translator with this complete plan.” Calvin continued sighing.
Now he finally knew why NASA’s researchers and engineers tolerated Lin Ran’s sharp tongue: the gap was too big, convinced in heart and mouth.
Especially counting co-occurrence frequency of adjacent vocabulary—not hard to think of, but they just didn’t.
Using statistical methods for ambiguity scenarios, adding statistical algorithms—that they hadn’t even dreamed of.
Dostert turned his head, smiled bitterly: “I’ve been pondering too. His statistical method sounds like fantasy, but the results are evident.
I estimate that under the professor’s guidance, the Georgetown translation system’s quality can improve by a large margin.
We won’t need carefully prepared short sentences; it can apply to broader scenarios, not limited to military domains.”
Calvin nodded: “Yeah, I didn’t believe it at first—language is clearly rule-driven; how can statistics solve it? But he shut me up with facts.
Worthy of the professor; his insight into the essence transcends fields.”
Dostert pondered: “You’re right; it feels like he can see through the essence of machine translation.
Maybe it’s the benefit of mathematical training; if I stay with the professor longer, I might want to pursue a PhD in mathematics.”
Calvin looked at him in surprise: “PhD in mathematics? Don’t joke.”
Dostert seriously said: “I’m not joking.
If mathematics can really help us better grasp the essence, pursuing a PhD in statistics isn’t impossible.”
Calvin was silent for a moment, then laughed: “If you go, I’ll go too.”
Dostert raised his beer glass almost overflowing: “Cheers to the professor! The professor will bring us victory!”
Calvin smiled back: “Cheers! But the professor’s temper—if only the professor were milder.”
On the other side, IBM’s two engineers Cuthbert Hurd and Peter Sheridan were utterly convinced by Lin Ran.
Cuthbert rubbed his temples, asked: “Peter, do you think the professor’s statistical model can really work?”
Peter put down his pen, smiled: “Cuthbert, to be honest, I was completely skeptical at first, but now I’m totally convinced. The professor’s method not only maximizes IBM 7090’s performance but also gives translation probabilistic support from chaos.”
Cuthbert nodded: “I see Georgetown University’s guys think so too; you didn’t see Calvin’s attitude shift from initial doubt to utmost seriousness.
The professor’s algorithm optimization is too perfect.”
Peter smiled bitterly: “Magic? As one of the era’s top mathematicians—maybe without ‘one of’—statistics is just a simple Sudoku game for the professor.
I just didn’t expect the professor to combine probability theory and linguistics so ingeniously; I never thought machine translation could be played this way.”
Cuthbert curiously asked: “You say the professor is fluent in Russian; those few Russian sentences today were standard beyond standard.
Spanning multiple fields too—no one in IBM, probably all of America, could come up with such a plan.
Could the professor be connected to the Soviet Union?”
Peter said speechlessly: “Would Soviet people let the professor stay in America?
If I were Nikita, I wouldn’t let talent like the professor stay in the White House.
Even gaining technical secrets from NASA, no amount of secrets could match the professor’s own value.
And have you thought: if the professor weren’t at NASA but in Moscow teaming with Korolev—imagine that scenario; could America win the space race?”
Cuthbert just thought about it and shook his head immediately: “Absolutely impossible.”
“So, if the professor had Soviet ties, how could he stay in America?
The first thing he’d do is lead manned moon landing in Moscow.” Peter laughed.
Scientists possibly colluding with Soviets is possible, but scientists with balance-altering influence—not likely.
If they just admired Lin Ran’s academic achievements, Watson admired him comprehensively.
Similar to John Morgan.
But Watson’s points of admiration differed from John Morgan’s.
“Professor, how did you think of building corporate image through an exhibition hall?” Watson raised his wine glass, smiling.
The Deep Blue exhibition hall by Times Square won IBM huge prestige.
Times Square had always been a New York landmark building, a must-checkpoint for almost every tourist to New York.
The Deep Blue exhibition hall attracted everyone’s attention with a style not of this era.
Combined with the world’s only artificial intelligence chess inside, capable of playing against humans automatically.
The shock to the public was unprecedented.
American companies have a long tradition of displaying technological strength and promoting their tech products through public exhibitions, traceable to the England period.
Whether Stephenson’s earliest trains or later ships, English people gathered the public and promoted lavishly in newspapers.
America’s earliest and most successful should be Edison’s light bulb; later Bell’s telephone was also a classic case.
But they were momentary; only that instant left an impression on the public—only when products entered daily life would they gain deeper recognition of the company and brand.
The Deep Blue exhibition hall’s existence left a profound impression of that Deep Blue and black-lined venue on every visitor.
IBM = artificial intelligence = high technology impression engraved in every visitor’s mind.
For IBM, it’s not just binding corporate image to artificial intelligence; it has nearly established that as long as the White House resolves to compete with the Soviet Union in artificial intelligence.
The supplier will have no other choice but IBM.
Equals Lin Ran’s suggestion giving IBM the world’s largest client out of thin air, with decades-long orders at minimum.
John Morgan’s General Aerospace got orders from NASA and at least gave Lin Ran shares; Watson gave nothing.
Forget Lin Ran implying he’s an idiot—even if Lin Ran pointed at his nose saying it, he’d just say yes yes yes, I am.
In the private room, the waiter quietly exited, leaving a quiet conversation space.
“Because I think artificial intelligence like Deep Blue should leave a sufficiently profound impression on the public.
Not just displayed internally at IBM.
As for why find artists for design—how could an ordinary theater match Deep Blue?”
Watson smiled and nodded: “You’re too right.
When I first saw Technology Ark completed, I felt it didn’t belong to this era; it was because of you that Deep Blue and Technology Ark were born.
Professor, I must toast you.”
Forget Hawking coming—he’d have to toast.
In front of Lin Ran, you Watson have to toast me.
Watson continued: “Professor, on behalf of IBM company, I extend our sincerest thanks.
Not just Deep Blue; your contributions to the Georgetown-IBM project are astounding. Your innovative methods will bring breakthrough progress to our machine translation system.”
Though not yet breakthrough progress, Watson was fully confident.
Lin Ran nodded: “That’s as it should be. Additionally, Watson, whether Deep Blue or Georgetown Translator, my contributions to IBM can’t be measured by money.”
Lin Ran was not humble at all.
This made Watson’s smile stiffen: “Professor, we’ll give you a sufficiently generous monetary reward.”
Lin Ran shook his head: “I’m not interested in money.”
Watson thought yes, he’d never heard the other was interested in money.
But asking for shares outright, Watson hesitated a bit.
“Professor.” Before Watson finished.
Lin Ran continued: “I need you to satisfy one small condition.
If you can’t meet my condition, I might have to seriously consider cooperating with General Electric.”
General Electric—key awareness.
Lin Ran’s relationship with the Morgan family goes without saying.
Watson knew it best.
General Electric also does computers.
Though General doesn’t do large-scale computers, General’s GE-225 series, as a transistor-based computer, is used for payroll, inventory management, accounting, etc.
General Electric has the capability and capital.
Plus Lin Ran and his master’s appeal, it could indeed threaten IBM massively.
Watson’s tone softened immediately: “Professor, what do you want?”
“MIT Radiation Laboratory Series”
