1
00:00:00,078 --> 00:00:04,444
[MUSIC]

2
00:00:04,444 --> 00:00:07,334
There are many systems for
restaurant reviews out there but

3
00:00:07,334 --> 00:00:11,118
today we're gonna talk about a pretty
new and exciting one that we can build.

4
00:00:11,118 --> 00:00:15,960
So it's an important day for

5
00:00:15,960 --> 00:00:21,100
us and we're trying to book
a Japanese restaurant, and

6
00:00:21,100 --> 00:00:26,120
we love Japanese food and want a really
cool place to have some Japanese food.

7
00:00:26,120 --> 00:00:29,670
But this is Seattle that we live in,
and Seattle is really awesome for

8
00:00:29,670 --> 00:00:33,526
Japanese food, so there are many,
many places for sushi, and in fact,

9
00:00:33,526 --> 00:00:38,500
there are many that have really
good ratings, say four stars.

10
00:00:38,500 --> 00:00:42,620
So perhaps when I'm thinking
about a restaurant to go to

11
00:00:42,620 --> 00:00:46,150
I'm not just thinking about the overall
stars of the restaurant, but

12
00:00:46,150 --> 00:00:50,440
I'm thinking about how is it in terms
of the food, in terms of the ambiance?

13
00:00:50,440 --> 00:00:53,090
In particular how is it
in terms of the sushi,

14
00:00:53,090 --> 00:00:54,580
which is the thing that
I really care about.

15
00:00:54,580 --> 00:00:58,863
I really love sushi and I want really,
really the best, freshest fish,

16
00:00:58,863 --> 00:01:01,290
and the most awesome,
innovative sushi I can get.

17
00:01:03,370 --> 00:01:07,360
So when you talk about a positive review
for a restaurant, and here's a restaurant

18
00:01:07,360 --> 00:01:12,670
that we love in Seattle, a sample
review might go something like this.

19
00:01:14,640 --> 00:01:17,600
So watching the chefs
create an incredible,

20
00:01:17,600 --> 00:01:20,810
edible piece of art,
was really a unique experience.

21
00:01:20,810 --> 00:01:26,060
So the experience is really positive if we
think about that aspect of the restaurant.

22
00:01:26,060 --> 00:01:30,040
But a review might want to say,
well, my wife tried the ramen, and

23
00:01:30,040 --> 00:01:31,810
it was pretty forgettable.

24
00:01:31,810 --> 00:01:33,990
That means the ramen, not so good.

25
00:01:35,560 --> 00:01:38,740
Now, the review continues and says,

26
00:01:38,740 --> 00:01:43,760
all the sushi was delicious,
easily the best sushi in Seattle.

27
00:01:43,760 --> 00:01:47,960
So the sushi was really two thumbs up.

28
00:01:47,960 --> 00:01:49,780
Now, I don't care about the ramen.

29
00:01:49,780 --> 00:01:51,410
I'm not going to this place for ramen.

30
00:01:51,410 --> 00:01:53,460
I don't care that this person
really didn't like the ramen,

31
00:01:53,460 --> 00:01:55,120
or his wife really didn't like the ramen.

32
00:01:55,120 --> 00:01:59,040
What I really care about is good
experience and amazing sushi.

33
00:01:59,040 --> 00:02:00,270
So this review is just for me.

34
00:02:01,780 --> 00:02:04,080
So when I'm thinking
about restaurant reviews,

35
00:02:04,080 --> 00:02:08,100
I wanna understand the aspects of
the restaurant, their positive or

36
00:02:08,100 --> 00:02:13,270
negative, and really think about which
ones of those really affect my interests.

37
00:02:13,270 --> 00:02:18,010
So if I look at all the restaurant
reviews, I want to feed them to this new,

38
00:02:18,010 --> 00:02:22,440
really exciting, new type of restaurant
recommendation system that's gonna tell me

39
00:02:22,440 --> 00:02:25,730
the experience was good, four stars.

40
00:02:25,730 --> 00:02:29,790
The ramen was a so-so, but who cares?

41
00:02:29,790 --> 00:02:34,560
But the sushi, the thing that I really
care about, that was five stars, and

42
00:02:34,560 --> 00:02:37,860
not just that, it's gonna give me
some interesting feedback, so for

43
00:02:37,860 --> 00:02:42,530
example, this is easily
the best sushi in Seattle.

44
00:02:42,530 --> 00:02:43,440
Now that's where I wanna go.

45
00:02:44,790 --> 00:02:48,890
So how do we build such an intelligent
restaurant review system?

46
00:02:48,890 --> 00:02:51,280
Well we're going to start
with all the reviews and

47
00:02:51,280 --> 00:02:53,580
we're going to break
them up into sentences.

48
00:02:53,580 --> 00:02:56,600
So each review is composed
of multiple sentences and

49
00:02:56,600 --> 00:03:00,510
some sentences cover different
aspects of the restaurant.

50
00:03:00,510 --> 00:03:05,240
So for example, a sentence that says
easily the best sushi in Seattle,

51
00:03:05,240 --> 00:03:08,520
I wanna feed that sentence
through a sentiment classifier.

52
00:03:08,520 --> 00:03:12,286
So sentiment classifier is gonna
say is this sentence positive, or

53
00:03:12,286 --> 00:03:13,814
is it a negative sentence?

54
00:03:15,717 --> 00:03:19,309
Now we have a sentiment classifier,
which can take a sentence and

55
00:03:19,309 --> 00:03:21,450
say is it positive or negative?

56
00:03:21,450 --> 00:03:25,088
So how do we build this new kind of
cool restaurant review experience?

57
00:03:25,088 --> 00:03:27,600
So we're gonna take all the reviews and

58
00:03:27,600 --> 00:03:29,980
break them into sentences
that will be discussed but

59
00:03:29,980 --> 00:03:33,970
then we're gonna select the sentences
that talk about the aspect I care about.

60
00:03:33,970 --> 00:03:36,600
So which ones of these
are really about sushi?

61
00:03:36,600 --> 00:03:38,538
So not all of them, but a subset of those.

62
00:03:38,538 --> 00:03:42,610
Then I feed those sentences
through the sentiment classifier,

63
00:03:42,610 --> 00:03:46,250
each one of them, and
then I average the results.

64
00:03:46,250 --> 00:03:50,330
In this case all of
the sentences were positive and

65
00:03:50,330 --> 00:03:56,390
so I'm gonna say this is a five star
restaurant review in terms of the sushi.

66
00:03:56,390 --> 00:04:00,720
Not just that, I can look at
the sentences about sushi and

67
00:04:00,720 --> 00:04:02,509
look at which ones are most positive.

68
00:04:03,570 --> 00:04:08,440
So for example, this one must be
the positive, but easily the best

69
00:04:08,440 --> 00:04:12,820
restaurant in Seattle maybe is the most
positive, and I can display those.

70
00:04:12,820 --> 00:04:15,470
I can also display the most negative ones.

71
00:04:15,470 --> 00:04:19,488
In this case, there's nothing
negative to say about the sushi.

72
00:04:19,488 --> 00:04:23,239
[MUSIC]