Summary of "PPDB: The Paraphrase Database" paper

PPDB.md

PPDB: The Paraphrase Database

Introduction

• The paper presents a database of ranked English and Spanish paraphrases derived by:
  • Extracting lexical, phrasal, and syntactic paraphrases from large bilingual parallel corpora.
  • Computing the similarity scores for the pair of paraphrases using Google ngrams and the Annotated Gigaword corpus.
• Link to the paper

Extracting Paraphrase from Bilingual Text

• The basic idea is that if two English strings e₁ and e₂ translate to the same foreign string f (also called pivot), they should have the same meaning.
• Informally speaking, the input to the system is translation triplets of the form < f, e, φ >, where
  • f is a foreign string
  • e is an english string
  • φ is a vector of feature functions
• The system can pivot over f to create paraphrase triplets < e₁, e₂, φ_p > where φ_p is computed using translation feature vectors φ₁ and φ₂
• For example, conditional paraphrase probability p(e₂|e₁) can be computed by marginalizing over all shared foreign language translations f:
  • p(e₂|e₁) = Sum over all f, p(e₂|f)p(e₁|f)

Scoring Paraphrases Using Monolingual Distributional Similarity

• Measure similarity of phrases using Distributional similarity.
• Can be used to rerank the paraphrases obtained from bilingual text or to obtain the paraphrases which could not be obtained from bilingual text alone.
• To describe a given phrase e₁, collect contextual features like:
  • n-gram based features for words (to the left and right of the given phrase)
  • Lexical, lemma-based, POS and named entity unigrams and bigrams
  • Dependency link features
  • Syntactic features
• Aggregate all the features, over all the occurences of e, to obtain distributional signature s_e.
• Define similarity between 2 phrases e₁ and e₂ as :
  • *sim(e₁, e₂) = dot(s_e1, s₂)/(|s_e1||s_e2|)
• Paper mentions two instances:
  • English paraphrases - 169.6 Million paraphrases
  • Spanish paraphrases - 161.6 Million paraphrases

Analysis

• The paper performed tests to analyse the precision-recall tradeoff for coverage of Propbank predictions and predicate-argument tuples.
• Human evaluation was performed over a sample of 1900 paraphrases to establish the correlation of PPDB scores with human judgement.

Areas of Improvement

• Segregation of data by domain or topic
• Support for more languages
• Improving paraphrasing scores by using additional sources of information and better handling of paraphrases ambiguity.

I would recommend contacting the authors.