Sentiment based Summarization of Restaurant Reviews

1
Sentiment based Summarization of Restaurant
Reviews

• Abhishek Gupta
• Tejaswi Tenneti
• Ankit Gupta

2
Which restaurant should we go to

• How often does this happen to you?

3
Input
4
Output
5
Roadmap

• Really glad that we have a great pizza place in
  the Heights. Pink's has tons of options and some
  great ingredients. Cramped place though.
• Feature Extraction
• pizza place, ingredients, place
• Feature Classification
• In food (ingredients), ambience (pizza place,
  place) and service
• Descriptor Extraction and Sentiment
  Classification
• great pizza place, great ingredients, cramped
  place

6
Data

• Crawl Yelp to get reviews of 120 restaurants

7
POS tagging and Parsing

• Generate POS tags of review sentences using
  Stanford POS tagger
• The/DT rude/JJ waiter/NN bought/VBD in/IN
  a/DT soggy/JJ pizza/NN
• Nouns in a sentence are features and
  adjectives are opinions
• Generate Dependency relations between words of a
  sentence using Stanford Parser.
• amod(rude, waiter) amod(soggy, pizza) .

8
Feature Extraction

• Features are words labeled NN, NNS and NNP
• Find frequent itemsets using Apriori Algorithm
• Eg Alto Palo pizza
• Ordering and Proximity Filtering
• Valid features in a sentence have words within a
  distance of 2.
• Standalone Identity Filtering
• Remove those features which occur mostly with
  their supersets.
• Eg pepperoni mostly comes with either topping,
  toppings or pizza.

9
Feature Extraction

• Infrequent Features are important!
• They have great lunch specials.
• Start with a list of all descriptors.
• For each sentence, look for a descriptor in the
  list.
• Use dependency tree to find NN or NNS it points
  to.
• Add such a word as infrequent feature.
• Capture a snippet of text around the feature.

10
Feature Classification

• Classify into food, ambience and service
• Training Data
• Features Hypernyms of a word
• pizza aliment, alimentation, dish, food,
  matter, nourishment, nutrient, nutriment,
  nutrition, physical entity, pizza, pizza pie,
  substance, sustenance, victuals

11
Feature Classification

• Use Multinomial Model over hypernyms of a word.
• Maximum Likelihood Estimation to calculate
  probability of a word being in a class.
• P(nutritionfood) words labeled food having
  nutrition / words labeled food.
• Added UNKNOWN word
• Used Absolute Discounting for smoothing counts.
• P(foodpizza)

12
Descriptor Extraction

• Use Stanford Parser to produce Dependency
  relations.
• nsubj(delicious, pizza), amod(meat, red)
• (The/DT pizza/NN was/VBD delicious/JJ)
  (Sam/NNP eats/VBZ red/JJ meat/NN)
• For each feature of a restaurant add the
  adjective as a descriptor

13
Sentiment Classification

• Classify descriptors into Positive, Negative,
  Neutral
• WordNet does not give the semantic orientation
  of adjectives
• Need to incorporate domain knowledge long wait,
  large slices
• Populate seed lists of adjectives. Generate
  synsets, related concepts, antonyms of a word and
  check for occurrence in the seed lists.
• 150 positive words, 80 negative words and 39
  neutral words

14
Results