Streaming Document Filtering using Distributed Non-Parametric Representations
Ignacio Cano, Sameer Singh, Carlos Guestrin

Citation
Ignacio Cano, Sameer Singh, Carlos Guestrin. "Streaming Document Filtering using Distributed Non-Parametric Representations". Talk or presentation, 30, October, 2014; Poster presented at the 2014 TerraSwarm Annual Meeting.

Abstract
When dealing with large, streaming datasets (such as of text documents), practitioners are often interested in identifying references to certain entities, and studying their prominence and topics over time. Current tools are quite restrictive for this setting; they are unable to handle streaming data, do not partition the references according to topics, and do not identify the vitality of the references. In this work, we propose a system that uses a flexible representation of entity contexts that are updated in a streaming fashion. We use distributed vector based representations of contexts, which address lexical sparsity in the data and generalize to unseen documents. Since each entity has several different aspects in which it may be mentioned (its topics), we propose a non-parametric clustering approach to identify which aspect of the entity is mentioned in each document. It is also crucial to capture temporal dynamics of each content; we propose a "staleness" measure to each entity topic, dynamically estimating the relevance of each entity. To enable analysis on massive scale streams, we propose updates to the topic identities, number of topics, and the staleness of topics in an online fashion, observing only a single document at a time. This combination of non-parametric clustering, staleness, and distributed word embeddings provides an efficient yet accurate representation of entity context that can be updated in a streaming manner. Our browser-based visualization demo provides an easy to use interface that enables users to switch between multiple entities of interest, select the time ranges to explore over, explore the prominence of topics over time, and understand the topics using word clouds.

Electronic downloads

Internal. This publication has been marked by the author for TerraSwarm-only distribution, so electronic downloads are not available without logging in.

Citation formats

• HTML

  Ignacio Cano, Sameer Singh, Carlos Guestrin. <a
  href="http://www.terraswarm.org/pubs/433.html"><i>Streaming
  Document Filtering using Distributed Non-Parametric
  Representations</i></a>, Talk or presentation, 
  30, October, 2014; Poster presented at the <a
  href="http://www.terraswarm.org/conferences/14/annual"
  >2014 TerraSwarm Annual Meeting</a>.

• Plain text

  Ignacio Cano, Sameer Singh, Carlos Guestrin. "Streaming
  Document Filtering using Distributed Non-Parametric
  Representations". Talk or presentation,  30, October,
  2014; Poster presented at the <a
  href="http://www.terraswarm.org/conferences/14/annual"
  >2014 TerraSwarm Annual Meeting</a>.

• BibTeX

  @presentation{CanoSinghGuestrin14_StreamingDocumentFilteringUsingDistributedNonParametric,
      author = {Ignacio Cano and Sameer Singh and Carlos Guestrin},
      title = {Streaming Document Filtering using Distributed
                Non-Parametric Representations},
      day = {30},
      month = {October},
      year = {2014},
      note = {Poster presented at the <a
                href="http://www.terraswarm.org/conferences/14/annual"
                >2014 TerraSwarm Annual Meeting</a>},
      abstract = {When dealing with large, streaming datasets (such
                as of text documents), practitioners are often
                interested in identifying references to certain
                entities, and studying their prominence and topics
                over time. Current tools are quite restrictive for
                this setting; they are unable to handle streaming
                data, do not partition the references according to
                topics, and do not identify the vitality of the
                references. In this work, we propose a system that
                uses a flexible representation of entity contexts
                that are updated in a streaming fashion. We use
                distributed vector based representations of
                contexts, which address lexical sparsity in the
                data and generalize to unseen documents. Since
                each entity has several different aspects in which
                it may be mentioned (its topics), we propose a
                non-parametric clustering approach to identify
                which aspect of the entity is mentioned in each
                document. It is also crucial to capture temporal
                dynamics of each content; we propose a "staleness"
                measure to each entity topic, dynamically
                estimating the relevance of each entity. To enable
                analysis on massive scale streams, we propose
                updates to the topic identities, number of topics,
                and the staleness of topics in an online fashion,
                observing only a single document at a time. This
                combination of non-parametric clustering,
                staleness, and distributed word embeddings
                provides an efficient yet accurate representation
                of entity context that can be updated in a
                streaming manner. Our browser-based visualization
                demo provides an easy to use interface that
                enables users to switch between multiple entities
                of interest, select the time ranges to explore
                over, explore the prominence of topics over time,
                and understand the topics using word clouds. },
      URL = {http://terraswarm.org/pubs/433.html}
  }
  

Posted by Ignacio Cano on 4 Nov 2014.