Entity-balanced Gaussian pLSA for Automated Comparison

Proceedings of NAACL-HLT 2016, pages 69–79, San Diego, California, June 12-17, 2016. c⃝2016 Association for Computational Linguistics Entity-balanced Gaussian pLSA for Automated Comparison Danish Contractor∗ IIT Delhi & IBM Research New Delhi, India dcontrac@in.ibm.com Mausam and Parag Singla IIT Delhi New Delhi, India {mausam,parags}@cse.iitd.ac.in Abstract Community created content (e.g., product de- scriptions, reviews) typically discusses one entity at a time and it can be hard as well as time consuming for a user to compare two or more entities. In response, we de- fine a novel task of automatically generating entity comparisons from text. Our output is a table that semantically clusters descriptive phrases about entities. Our clustering algo- rithm is a Gaussian extension of probabilis- tic latent semantic analysis (pLSA), in which each phrase is represented in word vector em- bedding space. In addition, our algorithm at- tempts to balance information about entities in each cluster to generate meaningful compar- ison tables, where possible. We test our sys- tem’s effectiveness on two domains, travel ar- ticles and movie reviews, and find that entity- balanced clusters are strongly preferred by users. 1 Introduction The proliferation of Web 2.0 has enabled ready ac- cess to large amounts of community created content, such as status messages, blogs, wikis, and reviews. These form an important source of knowledge in our day to day decision making, such as deciding which restaurant to try, or which movie to watch, or which city to visit etc. Unfortunately, such con- tent typically focuses on one real world entity at a time, whereas, a user deciding between alterna- tives is most interested in a comparative analysis of strengths and weaknesses of each. ∗This work was carried out as part of PhD research at IIT Delhi. The author is also a regular employee at IBM Research. There have been some recent attempts to create comparisons using expert knowledge, but generat- ing such comparisons manually does not scale – even pairwise comparisons are quadratic in the num- ber of entities. Few automated comparisons for spe- cific products with pre-defined attributes (e.g., lap- tops, cameras) exist; they are typically powered by existing structured knowledge bases. To the best of our knowledge, prior work on automatically gener- ating comparisons for arbitrary domains from un- structured text, does not exist. We define a novel task of generating entity com- parisons from textual corpora in which each docu- ment describes one entity at a time. For broad appli- cability, we do not restrict ourselves to a pre-defined ontology; instead, we use textual phrases that de- scribe entities as our unit of information. We call these descriptive phrases – they encompass general attribute-value phrases, opinion phrases, and other descriptions of the facets of an entity. We gener- ate entity comparisons in a tabular form where the phrases are organized semantically, thus, allowing for direct comparisons. Figure 1 shows a sample city comparison generated by our system for tourism. Our comparison generation algorithm extracts de- scriptive phrases per entity and clusters them into semantic groups. We …

Content truncated to 500 words