KG4IR - Workshop on Knowledge Graphs and Semantics for Text Retrieval and Analysis - SIGIR

A SIGIR tem um workshop específico para KG

 

 
The First Workshop on Knowledge Graphs and Semantics for Text Retrieval and Analysis
(KG4IR) - 2017

 

The availability of large knowledge graphs and semantic annotation techniques gave rise to successful approaches for many information retrieval (IR) tasks. It has been shown that heterogeneous information in knowledge graphs and entity annotations can help to significantly improve the performance of information retrieval tasks. 

Alignment includes the semantic annotation process such as entity linking of short keyword queries or relation extraction for satisfying information needs. It also includes information integration, ontology matching, entity search, and knowledge graph selection based on an information need.

 

Jun Xu, Chinese Academy of Sciences: Deep Approaches to Semantic Text Matching
In semantic matching the task is to predict relevance or similarity given two text documents. This task arises in many text applications such as paraphrase identification, information retrieval, and question answering. Focusing on neural network methods for semantic matching, Jun takes a new perspective on word-level matching and sentence-level matching. In both cases he applies word representations, proximity and multi-word patterns to bridging the semantic gap.

 

Hannah Bast, Universit ̈at Freiburg: Semantic Search on Text and Knowledge Bases
Hannah talks about how to craft systems that support search with meaning and gives an overview over her recent survey (Bast et al., 2016). While knowledge graphs are the preferred way of storing structured knowledge, she believes that most of the world’s knowledge will continue to be in text form. To support seamless search over both data sources, she presents two approaches: a system that supports a user in incrementally building a complex semi-structured query (such as the search in DBLP) and a system that allows free-form natural queries that are interpreted with respect to text and structured knowledge.

 

Hannah Bast, Bj ̈orn Buchhold, Elmar Haussmann, and others. 2016. Semantic Search on Text and Knowledge Bases. Foundations and Trends R© in Information Retrieval 10, 2-3(2016), 119–271. (Já li)

The Second Workshop on Knowledge Graphs and Semantics for Text Retrieval, Analysis, and Understanding (KG4IR) - 2018

 

Link -> https://kg4ir.github.io/

 

In particular, the semantics encoded in KGs are effectively integrated in query representation [5, 8, 11, 22], retrieval models [5, 15, 18], learning-to-rank [21], and generic representations [18]. Recent top-tier IR conferences feature many other papers on search systems are exploiting knowledge graphs [9, 13, 14, 19, 20, 23–25]. Some of these systems have also participated in TREC evaluations [16] or led to development of new TREC tracks [6]. Despite these successes, the utilization of knowledge graphs and semantics in information retrieval is still in its infancy and the community is actively working on how the different kinds of semantics in KGs can be utilized to improve end-to-end IR tasks.

Together we will work towards the ultimate goals of: (i) acquiring semantics for focused application
needs, (ii) aligning efforts around KGs and texts towards end-to-end usage, and (iii) building IR systems that more effectively utilize KGs, semantics, and alignment techniques.

 

RELATED WORKSHOPS

While a large number of workshops and tutorials on KGs are held in different venues, they focus either on KG construction, semantic search, or other applications; none of them have the focus on IR that we advocate. Our main focus lies on the use of KGs for retrieving, analyzing, and understanding text. By focusing on text as  the main target, our workshop differentiates itself from the general semantic search and knowledge graph topics. Nevertheless, we aim for a broadly inclusive community, encouraging participation from researchers with expertise in information extraction, semantic search, and text retrieval with the goal of promoting joint research.

 

Chenyan Xiong, Russell Power, and Jamie Callan. 2017. Explicit semantic ranking for academic search via knowledge graph embedding. In Proceedings WWW 2017. ACM, 1271–1279. (LER !!!)

Tutorial on Utilizing Knowledge Graphs for Text-centric Information Retrieval 

 

Furthermore, we are offering a SIGIR conference tutorial on foundations and best practices on utilizing
knowledge graphs for text-centric information retrieval. The goal is to jump-start any information retrieval researcher who is interested in contributing to this line of work in the future.

 

The depth and breadth of content in these KGs made them not only rich sources of structured knowledge by themselves, but also valuable resources for search systems.

 

We use the term entity to denote any entry in a KG, while distinguishing it from a mention of an entity in text (which was previously referred to as entity in the literature about named entity recognition). As such, we leverage a generic, extended definition of entities to encompass any entry in a KG, which may include, for example, people and places, but also chemical compounds, diseases, as well as intangible concepts, such as “information retrieval”.  

Entity retrieval, then, refers to the task of retrieving relevant KG entries in response to a user query. 

 

Entity linking refers to the annotation of text such that all entity mentions are annotated with identifiers to KG entries. Accurate entity linking methods play a critical role in this scenario, as they provide a bridge between unstructred text and structured information about entities in KGs.

 

We also touch on the issue of semantic search by providing an overview of novel and recent advances in entity retrieval that are not covered in previous tutorials on this topic. This tutorial focuses on the use of KGs for text-centric information retrieval and, more specifically, on how to leverage different types of data provided by KGs for ad hoc document retrieval and other search systems

 

While each KG has unique characteristics, KG entities are typically associated with different names and,  possibly, types from a taxonomy or category system, as well as have relations with other entities.Some KGs also incorporate explicit textual descriptions for entities and/or links to textual documents that are “about” each entity. 

 

Throughout the tutorial, we discuss how each of these different types of information can be used to: a) retrieve a set of entities for an information need formulated as a keyword query or a question, or more broadly: how to assess the relevance of KB elements to a given topic, b) how to recognize mentions of entities from a KG in a textual fragment and c) how to utilize these mentions to assess the relevance of a textual fragment.

 

Entity linking [48] is the task of identifying entity mentions in text and aligning them with their corresponding entities in the knowledge graph. Entity linking systems are typically structured as a pipeline. The first step is to identify linkable phrases, i.e., text segments that could mention an entity. In the second step, a candidate set for each such phrase is retrieved, of course the possibility that the mentioned entity is not contained in the knowledge graph (so-called NIL entities) must be considered. The final step is to disambiguate which of the candidate entities are actually referred in the mention based on the context of the mention. (A desambiguação depende do contexto/tema/assunto)

 

Research on vertical, composite, and aggregate search provides an alternative perspective on the problem, where the main task is to combine information from various resources. 

 

KG-aware document retrieval models incorporate matches of entity names, contextual terms, and entity links. Together with approaches for finding relevant entities these give rise to an effective generalizable retrieval approach. Different machine learning approaches aid in solving this task. Concept Feedback [33] uses a feature-based system with graph walks. Latent Entity Space [38] uses generative language models.
EsdRank [55] and Entity Query Feature Expansion [12] integrate entity retrieval, text retrieval, and different indicators from KG-based query expansion with a supervised learning-to-rank approach.

 

Research on diversification relies on the identification of different subtopics within query-relevant material. Entity-centric approaches can be applied to topic detection [46]. Especially for complex information needs, it becomes more important to organize topics for coherent presentation [3, 16]. It seems sensible that knowledge
graphs can help here, but more work on utilizing KGs in the identification of query sub-topics is needed.