Study in IRLAB

Text Semantic Matching Review

Posted on | In

Text Semantic Matching Review

Several papers about Text Semantic Matching

Short Text Similarity with Word Embeddings

a model for short text similarity using word embeddings created by Tom Kenter and Maarten from University of Amsterdam for CIKM 2015

Abstract

  • Keywords: Short Text Similarity, Word Embeddings
  • Backgroud: Determining semantic similarity between texts is important in many tasks in information retrieval such as search, query suggestion, automatic summarization and image finding.
  • Previous work:
    • Methods: many approaches have been suggested, based on lexical matching, handcrafted patterns, syntactic parse trees, external sources of structured semantic knowledge and distributional semantic.
    • Drawbacks:
      • lexical features: like string matching, do not capture semantic similarity beyond a trivial level.
      • handcrafted patterns and external sources of structured semantic knowledge: can’t be assumed to be available in all circumstances and for all domains
      • approaches dependings on parse trees are restricted to syntactically well-formed text, typically of on sentence in length.
  • This work:
    • Try: determining short text similarity is possible using only semantic features.
    • using word embeddings to pertain to a representation of meaning.
    • They show on a publicly available evaluation set commonly used for the task of semantic similarity that their method outperforms baseline methods that work under the same conditions.
  • Distributional semantics: LSA, word embeddings, GloVe
  • Text-level semantics without external semantic knowledge: CNN, et al.
  • Text-level semantics with external knowledge: using sources of structured semantic knowledge like Wikipedia and WordNet for semantic text similarity task.
  • SemEval STS: two evaluation for Semantic Text Similarity task.
  • Meta-level features:

Short Text Similarity With Semantics Only

The meaning of longer pieces of text(containing multiple terms) can be captured by taking the mean of the individual term vectors.

Question Similarity Modeling with Bidirectional Long Short-Term Memory Neural Network

BiLSTM Memory Neural Network for question similarity modeling created by Chao An and Jiuming Huang from NUDT for IEEE First International Conference on Data Science(I don’t know what it is)

Abstract:

  • Background: Modeling sentence similarity all along is a challengeable task in the field of natural language processing(NLP), since ambiguity and variability of linguistic expression.
  • This work:
    • Methods: Bidirectional Long Short-Term Memory(BLSTM) neural networks, suited for sentence similarity computation, paraphrase detection, question answering and so on.
    • Dataset: labeled Yahoo! Answers data
    • Note: this model doesn’t use external resources, such as WordNet or parsers.

Bilateral Multi-Perspective Matching for Natural Language Sentences

a model for sentence matching created by Zhiguo Wang, Wael Hamza, Radu Florian from IBM Watson Research Center published on arXiv

Abstract:

  • Backgroud: Natural language sentence matching is a fundamental technology for a variety of tasks.
  • Previous work:
    • match sentences from a single direction or only apply single granular(word-by-word or sentence-by-sentence) matching.
  • This work:
    • Model: bilateral multi-persective matching(BiMPM) model: an inner matching model
    • evaluate tasks: paraphrase identification, natural language inference and answer sentence selection.
    • Experimental results on standard benchmark datasets show that our model achieves the state-of-the-art performance on all tasks.

  • Siamese architecture: CNN or RNN is applied to two input sentences individually, and a matching decision is made solely based on the two sentence vectors.

    • advantage: sharing parameters makes the model smaller and easier to train.
    • disadvantage: there is no interaction between the two sentences during the encoding procedure, which may lose some important information.

  • Matching aggregation framework:

A Deep Relevance Matching Model for Ad-hoc Retrieval

another deep matching model created by Jiafeng Guo, Yixing Fan, Qingyao Ai and W. Bruce Croft from CAS for CIKM 2016

Abstract:

  • Background:
    • deep neural networks have led to exciting breakthroughs in speech recognition, computer vision, and natural language processing(NLP) taks but ad-hoc retrieval.
    • a matching problem between two pieces of text in existing work using deep models, and treated equivalent to many NLP tasks such as paraphrase identification, question answering and automatic conversation.
  • Differences between relevance matching and semantic matching:
    • exact matching signals
    • query term importance
    • diverse matching requirements
  • This work:
    • They propose a novel deep relevance matching model(DRMM) for ad-hoc retrieval: a joint deep architecture at the query term level for relevance matching, using matching histogram mapping, a feed forward matching network, and a term gating network.
    • Experimental results on two representative benchmark collections show that our model can significantly outperform some well-known retrieval models as well as states-of-the-art deep matching models.