The reason for such an odd result in the previous section can be traced back to the context.  Notice that review 3 had the phrase No nonsense, no gimmicks, which is largely positive, but has two negative words attached to it. How can we take context into account? Enter n-grams. An n-gram is a sequence of n consecutive items (words, or in the case of speech, phonemes) from a given sequence of text or speech. Let's make this clear with an example and use 2-grams, or bigrams:

text_df <- data_frame(line = 1:4, text = text)
text_df <- text_df %>%
 unnest_tokens(bigram, text, token="ngrams", n=2)
text_df

This gives us the following:

# A tibble: 70 x 2
 line bigram
 <int> <chr>
 1 1 the food
 2 1 food is
 3 1 is typical
 4 1 typical czech
 5 1 czech and
 6 1 and the
 7 1 the beer
 8 1 beer is
 9 1 is good
10 1 good the
# ... with 60 more rows

So, we now see that consecutive words are put together. Already, this can be helpful enough to determine the negations of negative words that are actually positive, as in our preceding review 3. Let's find out which negative words are negated. First, we split the two words of the bigram into two columns:

library(tidyr)
text_df <- text_df %>% separate(bigram, c("w1","w2"), sep=" ")
text_df

Which gives us:

# A tibble: 70 x 3
 line w1 w2
 * <int> <chr> <chr>
 1 1 the food
 2 1 food is
 3 1 is typical
 4 1 typical czech
 5 1 czech and
 6 1 and the
 7 1 the beer
 8 1 beer is
 9 1 is good
10 1 good the
# ... with 60 more rows

A bit of dplyr magic brings the offending part of the sentence to question:

text_df %>% 
 filter(w1=="no") %>% 
 inner_join(afinn, by=c(w2="word"))

Which is:

# A tibble: 1 x 4
 line w1 w2 score
 <int> <chr> <chr> <int>
1 3 no nonsense -2

We could now use this information to override the score of our third review. However, note how involved a process it might be. Sure, in many cases it would work well, but we should find more systematic ways of dealing with context.