Download presentation
Presentation is loading. Please wait.
Published byΘεράπων Ζυγομαλάς Modified over 5 years ago
1
An Approach to Abstractive Multi-Entity Summarization
龚赛赛
2
Contents Background and Motivation Related Work Method Framework
3
Background Multi-entity summarization: distill the major information from a set of entities to generate a compressed summary Useful for many tasks, including entity search and entity browsing Potential for QA (cp. multi-document summarization for QA)
4
Motivation Existing works mainly generate extractive summaries for an entity set, usually an entity pair, to find certain similarities, differences or identity information Extract a subset of sentences/features/property-value pairs for different entities Not efficient for identifying similarities and differences for a set of more than 2 entities
5
Motivation Example Better summary using count
All the 23 persons affiliation Websoft 3 status teachers, 5 status phdstudents, … 15 of 23 from jiangsu 3 deliver www paper … …… Affiliation:websoft …… Affiliation:websoft Affiliation:websoft Affiliation:websoft status: Professor status : Professor status: AP status: phd student status: phd student …… …… From: jiangsu From: jiangsu From: jiangsu From: jiangsu From: jiangsu Deliver www paper Deliver www paper Deliver www paper
6
Motivation Generate abstractive summary for finding commonality (similarity) of the entire set and particularity(difference) of each entity Using count Using subclass,subproperty reasoning Using literal value partitioning
7
Summary UI Commonality panel Difference panel (focus on difference) …
All the 23 persons affiliation Websoft 3 status teachers, 5 status phdstudents, … 15 of 23 from jiangsu … Deliver 2 www papers Leader
8
Related work Single entity summarization ([1,3])
Find important information Based on centrality, relatedness, informativeness, uniqueness and so on Consider diversity using linear programming, MMR or clustering Entity pair summarization ([2]) Identify similarities and differences based on uniqueness and property-value similarity/dissimilarity Consider diversity Entity set ([4]) Identify uniqueness, differences and contextual relatedness for facilitating entity linking Structured summary Path for entity pairs [6], Steiner tree for input entities [5], connected subgraph for input entities [7] Not condensable
9
Method Framework Two summaries: Commonality summary for identifying the major common information of the set (entire vs subset), and single summary for identifying the particularity of a single entity Single summary: extract top-K property-value pairs Informal commonality summary statement ? Given the entity set with their descriptions, generate top-K “facets”? Facets allowing entity num count, literal value interval ? A facet consists of a property with its value combination? 20 students. 5 are of phd students. 9 are of msc students. 5 persons age persons age 28+ .
10
Commonality Summary Value combination e.g.
For types, get superclasses. Select suitable classes to group/partition entities For literals, find suitable intervals to group/partition entities Handle equivalent/synonymous props Replace to a representative one Basic idea for generating top-k facets Covered entity number The group number of entities inducing from the facet Property importance
11
Single Summary Remove property-value pairs that are similar to the facets in commonality summary Extract property-value pairs as summary based on informativeness and uniqueness.
12
Evaluation Plan Entity set: an entity with its neighbors
Invite expert to make guided summary Compare automatic summary with reference
13
Evaluation Plan
14
Reference FACES: Diversity-Aware Entity Summarization using Incremental Hierarchical Conceptual Clustering. aaai Facilitating human intervention in coreference resolution with comparative entity summaries. eswc Relin: relatedness and informativeness-based centrality for entity summarization. iswc Summarizing Entity Descriptions for Effective and Efficient Human-centered Entity Linking. www STAR: Steiner Tree Approximation in Relationship-Graphs Rex:Explaining relationships between entity pairs Fast discovery of connection subgraphs
Similar presentations
© 2024 SlidePlayer.com. Inc.
All rights reserved.