4 Filip Ilievski, Pedro Szekely, and Bin Zhang
ConceptNet. Nevertheless, these high-quality mappings provide an opening for
consolidation of commonsense knowledge, a goal we pursue in this paper.
2.2 Challenges
Combining these sources in a single graph faces three key challenges:
1. The sources follow different knowledge modeling approaches. One such
difference concerns the relation set: there are very few relations in ConceptNet
and WordNet, but (tens of) thousands of them in Wikidata and Visual Genome.
Consolidation requires a global decision on how to model the relations. The gran-
ularity of knowledge is another factor of variance. While regular RDF triples fit
some sources (e.g., ConceptNet), representing entire frames (e.g., in FrameNet),
event conditions (e.g., in ATOMIC), or compositional image data (e.g., Visual
Genome) might benefit from a more open format. An ideal representation would
support the entire granularity spectrum.
2. As a number of these sources have been created to support natural language
applications, they often contain imprecise descriptions. Natural language
phrases are often the main node types in the provided knowledge sources, which
provides the benefit of easier access for natural language algorithms, but it intro-
duces ambiguity which might be undesired from a formal semantics perspective.
An ideal representation would consolidate various phrasings that share a concept
or a referent, while still allowing easy and efficient access to these concepts based
on their natural language labels or aliases.
3. Although these sources contain links to existing ones, we observe sparse
overlap. As these external links are typically to WordNet, and vary in terms
of their version (3.0 or 3.1) or target (lemma or synset), the sources are still
disjoint and establishing (identity) connections is difficult. Bridging these gaps,
through optimally leveraging existing links, or extending them with additional
ones automatically, is a modeling and integration challenge.
Previous efforts that combine commonsense resources exist. A unidirectional
manual mapping from VerbNet classes to WordNet and FrameNet is provided
by the Unified Verb Index [31]. The Predicate Matrix [6] has a full automatic
mapping between lexical resources, including FrameNet, WordNet, and Verb-
Net. PreMOn [5] formalizes these in RDF. In [19], the authors produce partial
mappings between WordNet and Wikipedia/DBpedia. BabelNet [21] integrates
several commonsense sources, but extending BabelNet is prevented by its license.
Recent systems integrate parts of these sources in an ad-hoc manner to reason
on a downstream task, e.g., [35] combine edges from Visual Genome, WordNet,
and ConceptNet in a neural network that produces a scene graph from an image.
3 The Common Sense Knowledge Graph
3.1 Principles
To address the aforementioned challenges, we devise principles and a respective
representation format that are driven by pragmatic goals of simplicity, modular-
ity, and utility. Specifically, it should be simple to integrate the graph represented
剩余15页未读，继续阅读
syp_net
- 粉丝: 158
- 资源: 1187
我的内容管理
展开
最新资源
- 岩石滑动与断层冲击地压：声发射特征分析
- 新时代煤炭工业八大战略新取向剖析
- 开采强度对华亭矿区冲击地压危险性的影响分析
- 新安煤矿探采对比：影响安全生产的关键因素与技术支撑
- 210MW循环流化床锅炉引风机变频改造优化与节能效果
- 浅埋煤层开采沉陷预测：方法应用与影响分析
- 优化巷道过陷落柱支护方案：数值模拟与现场试验
- 坚硬顶板定向断裂爆破技术：安全高效采煤新方案
- 北航2019算法课作业解析：投资策略与凸子多边形优化
- 断层破碎带垮冒下堆积体边界抗力分布关键影响因素
- 哈拉沟选煤厂优化末煤系统提高效率与处理能力
- 瓦斯涌出影响下采动煤体蠕变失稳研究
- 提升煤矿装载效率的可控箕斗定量装载设备设计
- 水分调控下烟煤瓦斯吸附特性对比研究
- 重介质悬浮液密度控制硬件设计与选型探讨
- 坚硬顶板巷道支护参数设计：现场测试与数值分析
资源上传下载、课程学习等过程中有任何疑问或建议，欢迎提出宝贵意见哦~我们会及时处理！
点击此处反馈
