跟讀練習: The Math Behind Recommender Systems - 通過影片學習英語口語

選擇練習模式

每種模式都會開啟全屏練習室。

89 句
如果句子過短或過長,請點選 Edit 進行調整。
1
In 2006, Netflix famously offered a prize to solve a simple problem which had been around for years.
2
All online commerce giants are fighting to get better at solving this same problem.
3
Given observations of a user's past behavior, predict which other things that same user will like.
4
For example, based on Alice's viewing history, will she like
5
The Matrix or Grapes of Wrath?
6
We can represent user preferences graphically as connections between people on one side and things on the other, such as movies.
7
Here, the strength of connection is visually represented by the thickness of the lines.
8
So the problem is to fill out all of these missing connections we don't yet know, shown in red.
9
But because of the messy, unpredictable nature of people's taste, we can never perfectly predict what they will like
10
because for one it involves a guess about the future based on something you've never seen
11
and two the answer is always in flux
12
because people's tastes change over time what we can do is try to estimate those values as best
13
we can using whatever data we have access to
14
so regardless of the context we can describe the problem in the following mathematical form we use a matrix
15
which is a collection of numbers organized into rows and columns, where the rows are users and the columns are items, in this case movies.
16
And each cell in this matrix contains a number that describes how much each individual likes each movie.
17
For example, we could use a scale from 0 to 4, where 0 would represent I hate it,
18
2 is neutral, and 4 is I love it.
19
At any given moment, there will be some set of preference data for users and movies.
20
But it will be incomplete because users have only seen and rated a few movies each.
21
And so the problem boils down to trying to predict these missing values.
22
How should we make these predictions?
23
There are two general approaches to this problem.
24
The first is called content filtering, which is to use information we know about people and things as connective tissue for recommendations.
25
It's like asking Alice questions about what kinds of movies she likes to figure out if she will like The Matrix.
26
It begins by labeling each person and movies with some known attributes, or what we call features, for example, action and comedy.
27
That means if Alice likes comedy and hates action, we can represent her as 3 -0.
28
Each movie is also mapped to each feature in the same way.
29
For example, the matrix has no comedy and lots of action, so we could represent it as 0, 4.
30
And to determine whether someone will like a movie, we need to multiply these factors together.
31
So we could represent the strength of connection between Alice
32
and the matrix as 3 times 0 plus 0 times 4 equals 0.
33
So our estimation is that she will hate the movie.
34
So to make our predictions, we first need to gather this feature data for every user and movie.
35
To simplify, let's return to a matrix representation.
36
We can store this information in two matrices.
37
One defines the mapping between people and features.
38
The other defines the mapping between movies and features.
39
And by multiplying these two matrices together, using matrix multiplication, we get an estimated strength of connection between every person and movie.
40
And if we'd like, we can normalize our data to make sure the scale stays between 0
41
and 4 by dividing all values by 8 and rounding to the nearest 0 .5.
42
That's one way to solve this problem, known as content filtering.
43
The problem is, it's overly simplistic and not very accurate it.
44
That's because there are obviously more relevant features of a movie than just comedy or action.
45
The obvious way to improve this is just to include more features.
46
If you recall when Netflix started, it did just this.
47
It would ask new users to fill out a laundry list of questions about their preferences before presenting them with suggested movies.
48
And this leads to the problem of having to collect all of this preference data on users.
49
Not only is it a burden for users, it's also prone to failure, as we aren't always great at describing our own preferences.
50
Sometimes we simply can't explain why we like things.
51
We just do.
52
Which brings us to the other approach, called collaborative filtering.
53
The motivation for collaborative filtering comes from the idea that you will probably like things people with similar viewing habits also like.
54
This idea was popularized in this 2009 paper by Cohen, Bell, and Walensky.
55
So to begin, we can throw away the idea of dreaming up features used to connect peoples in movies.
56
Instead, we flip things around and use the user preference data we do have to generate the features.
57
For example, we might have this incomplete set of preference data,
58
and we will instead learn or discover the relevant features based on patterns in this data.
59
And this is done by simply reversing the problem.
60
We first perform an approximate factorization into two matrices, and we can do this using a machine learning approach.
61
The job of the machine learning algorithm is to guess values for those matrices,
62
which will match the existing data in the preference matrix as closely as possible.
63
The simplest approach is to simply guess numbers over
64
and over until you arrive at a set of numbers which predict the data with the lowest error overall.
65
Once this estimation is finished, we can multiply the matrices as before to fill in all of the missing values.
66
And it's important to note that we won't know exactly what to label these discovered features, so we call them latent features,
67
because they arise out of the underlying patterns in the data.
68
You can think of them as an average or weighted sum of the patterns in the data.
69
They are not based on a human -defined feature, such as comedy.
70
And that's the key insight behind this method.
71
With content filtering, the features come from the human mind,
72
whereas with collaborative filtering, the features are extracted directly from the patterns in the data.
73
And this will predict the data in the same way but more accurately.
74
It's also important to notice that this kind of prediction is actually a form of compression
75
because we can represent a huge matrix of preference data by two smaller matrices of feature data.
76
And once our feature data is determined or learned, we can throw away the original data we started with,
77
because we can simply multiply the features together to recreate that data.
78
And the reason this is possible is because the preference data we started with is not random, but follow patterns which exist across people.
79
If instead the preference data was filled with random numbers,
80
we would not be able to compress it accurately because there would be no patterns to extract.
81
This collaborative filtering approach has also been applied to a wide range of contexts,
82
such as evaluating policy outcomes before they are implemented in the real world, known as synthetic control.
83
For example, if we wanted to understand the effect of gun control or minimum wage increases in some city,
84
we could look at the effect in similar cities first
85
which have implemented those policies to approximate the outcome using the same mechanism.
86
So it is the patterns we share that allow this method to work.
87
Recommendation systems based on these ideas are in widespread use today, whether it's movies, music, news articles, or anything else you search for online.
88
The things
89
that are recommended to you are based on patterns the machine has observed in other people who are similar to yourself.

下載應用

流利口語所需的一切

AI 發音評分逐句打分
音標練習掌握每個音
詞彙學習建立你的詞庫
詞彙遊戲邊玩邊學
FOR YOU

推薦影片

本課的詞彙與口語要點

這節口語課以影片“The Math Behind Recommender Systems”為素材。 影片中反覆出現的詞有:feature, movie, preference, matrix, user。 這段影片共有 89 個句子、1284 個單詞可供跟讀。 講話部分時長為 8:04。 說話人語速自然,每分鐘約 159 個詞,接近日常對話。 84% 的單詞屬於英語最常用的 3,000 詞,其餘的詞建議在練習前先學一下。

影片中的重點詞彙

影片中 15 個值得學習的單詞,附發音和釋義:

單詞發音釋義
preference 名詞/ˈpɹɛf(ə)ɹ(ə)ns/優選 /優選, 選擇 /選擇
filter 名詞/ˈfɪltɚ/過濾器 /過濾器, 濾過器 /濾過器
predict 動詞/pɹɪˈdɪkt/預言 /預言
collaborative 形容詞/kəˈlæb.ə.ɹə.tɪv/協作的 /協作的
solve 動詞/sɒlv/解決 /解決
define 動詞/dɪˈfaɪn/界定
prediction 名詞/pɹɪˈdɪkʃən/預言 /預言, 預估 /預估
label 名詞/ˈleɪ.bəl/標籤 /標籤, 簽條 /籤條
discover 動詞/dɪˈskʌvə/發現 /發現, 發覺 /發覺
estimate 名詞/ˈɛs.tɪ.mət/估計 /估計
implement 動詞/ˈɪmpləmənt/實施 /實施, 執行 /執行
recommendation 名詞/ˌɹɛkəmɛnˈdeɪʃən/建議 /建議, 勸告 /勸告
extract 名詞/ˈɛkstɹækt/精
incomplete 形容詞/ɪn.kəmˈpliːt/不完整的, 不完全的
wage 名詞/weɪd͡ʒ/工資 /工資, 薪金

影片中出現的短語動詞

單詞發音釋義
fill out 動詞/fɪl ˈaʊt/填, 填寫 /填寫
throw away 動詞棄置 /棄置, 扔
figure out 動詞弄清楚

值得反覆跟讀的句子

影片中簡短而完整的句子,可以直接用在日常對話裡:

  • How should we make these predictions?
  • To simplify, let's return to a matrix representation.
  • Sometimes we simply can't explain why we like things.

影片中的語法

說話人最常用的結構,並附上影片中的原話:

結構影片中的用法
被動語態 be + 過去分詞 — 強調發生了什麼,而不是誰做的is visually represented · is called · is also mapped
現在完成時 have/has + 過去分詞 — 過去發生但與現在仍有關聯的事you've never seen · have only seen · has also been applied
定語從句 who / which + 從句 — 補充說明人或事物filtering, which is · people who are

需要注意的發音

說話人用了 6 次縮略和弱讀形式,例如 aren't, can't, don't。請按聽到的簡短形式來說。

  • “th” 音: gather /ˈɡæðɚ/, mathematical /ˌmæθ(.ə)ˈmæt.ɪ.kəl/, algorithm /ˈælɡəɹɪðm̩/, synthetic /sɪnˈθɛtɪk/, thickness /ˈθɪknəs/
  • “sh” 和 “zh” 音: prediction /pɹɪˈdɪkʃən/, recommendation /ˌɹɛkəmɛnˈdeɪʃən/, representation /ˌɹɛp.ɹə.zɛnˈteɪ.ʃən/, tissue /ˈtɪʃ.(j)u/, motivation /ˈmoʊ.tə.veɪ.ʃən/
  • 長單詞——注意重音位置: collaborative /kəˈlæb.ə.ɹə.tɪv/, recommendation /ˌɹɛkəmɛnˈdeɪʃən/, accurately /ˈæk.jə.ɹɪt.li/, approximate /əˈpɹɑk.sə.mət/, representation /ˌɹɛp.ɹə.zɛnˈteɪ.ʃən/

中文母語者容易讀錯的音:

  • /v/ — 不要讀成 /w/,上齒輕觸下唇: collaborative /kəˈlæb.ə.ɹə.tɪv/, solve /sɒlv/, discover /dɪˈskʌvə/, arrive /əˈɹaɪv/, reverse /ɹɪˈvɝs/
  • 詞尾子音 — 要讀清楚,後面不要加母音: matrix /ˈmeɪ.tɹɪks/, preference /ˈpɹɛf(ə)ɹ(ə)ns/, predict /pɹɪˈdɪkt/, collaborative /kəˈlæb.ə.ɹə.tɪv/, solve /sɒlv/
  • /r/ 和 /l/ 的區別: collaborative /kəˈlæb.ə.ɹə.tɪv/, accurately /ˈæk.jə.ɹɪt.li/, regardless /ɹɪˈɡɑɹd.lɪs/, recall /ɹɪˈkɔl/, neutral /ˈnuː.tɹəl/

如何用這段影片練習

  1. 先完整聽一遍影片,不要開口,記下不認識的單詞。
  2. 先用 0.75 倍速逐句跟讀,熟練之後再回到正常速度。
  3. 錄下自己的聲音並與原聲對比,特別注意 preference, filter, predict 這類單詞。

什麼是跟讀法?

跟讀法 (Shadowing) 是一種有科學依據的語言學習技巧,最初開發用於專業口譯員的培訓,並由多語言者Alexander Arguelles博士普及。這個方法簡單而強大:您在聽英語母語原聲的同時立即大聲重複——就像是一個延遲1-2秒緊跟說話者的影子。與被動聽力或語法練習不同,跟讀法強迫您的大腦和口腔肌肉同時處理並模模擬實的講話模式。研究表明它能顯著提高發音準確性,語調,節奏,連讀,聽力理解和口語流利度——使其成為雅思口語備考和真實英語交流最有效的方法之一。

影子跟讀法: 閱讀完整分步指南 →