Non-monotonic Generation of Knowledge Paths for Context Understanding
Abstract
1 Introduction
1.1 Background
1.2 Problem Definition
1.3 Research Objectives and Contributions
2 Related Works
2.1 Knowledge-enhanced Information Systems
2.2 Knowledge Reasoning
2.3 Conditional Sequential Generation
3 Proposed Framework and CPG Models
3.1 Two-Stage Framework
3.2 Context Extractor
3.2.1 KEMatch: Knowledge-enabled Embedding Matching.
3.2.2 LTRMHSA.
3.3 CPG
3.3.1 Pretraining of Non-monotonic Path Generation Model.
3.3.2 Fine-tuning Non-monotonic CPG.
3.4 Tree Serialization
4 Dataset and Evaluation Metrics
4.1 Collection of Wikinews Datasets
Wikinews Dataset | ||
---|---|---|
Wiki-film | Wiki-music | |
# context documents | 40 | 40 |
# entities mentioned | 563 | 471 |
# \((e_h,e_t)\) entity pairs | 1,396 | 1,237 |
Knowledge Graph | ||
# entities | 59,173 | 44,886 |
# relations | 651 | 513 |
Ground Truth Contextual Paths | ||
Avg./Max. path length | 3.87/6 | 3.62/6 |
# unique path entities | 563 | 471 |
# unique path relations | 139 | 108 |
4.2 Evaluation Metrics
5 Experiment Results
5.1 Effectiveness of Context Extractor Methods
Context Extractor | Feature Combinations | Precall | Recall | |
---|---|---|---|---|
Context Window | 68.4 | 62.3 | ||
KEMatch | AVG Entity Rep. | + | 73.5 | 70.9 |
\(\odot\) | 75.6 | 73.8 | ||
\(-\) | 72.9 | 70.1 | ||
all | 72.1 | 69.3 | ||
Mention Paragraph Rep. | + | 75.4 | 73.2 | |
\(\odot\) | 80.1 | 77.6 | ||
\(-\) | 72.7 | 71.6 | ||
all | 71.5 | 69.8 | ||
Title + Mention Paragraph Rep. | + | 69.20 | 65.4 | |
Cxt. Query Entity Rep. | + | 74.7 | 72.3 | |
\(\odot\) | 81.3 | 78.9 | ||
\(-\) | 72.5 | 71.7 | ||
all | 71.9 | 70.4 | ||
LTRMHSA | 85.9 | 83.1 |
5.2 Performance in CPG
Dataset | Wiki-Film | Wiki-Music | ||||||
---|---|---|---|---|---|---|---|---|
Feat. Comb. | %Recov | AVG PW Sim | NGEO(E), NGEO(R) | %Recov | AVG PW Sim | NGEO(E), NGEO(R) | ||
NCNMPG | 19.7 | 0.44 | 0.29, 0.24 | 20.32 | 0.46 | 0.29, 0.23 | ||
NMCPGT | Random Context | 62.33 | 0.59 | 0.26, 0.22 | 60.15 | 0.58 | 0.27, 0.22 | |
Context Window | 73.27 | 0.75 | 0.2, 0.19 | 75.22 | 0.74 | 0.21, 0.2 | ||
KEMatch * AVG Ent Rep. | + | 76.76 | 0.83 | 0.17, 0.16 | 78.13 | 0.81 | 0.17, 0.15 | |
\(\odot\) | 78.14 | 0.84 | 0.17, 0.16 | 78.92 | 0.81 | 0.17, 0.15 | ||
\(-\) | 71.11 | 0.78 | 0.19, 0.18 | 70.37 | 0.80 | 0.2, 0.18 | ||
all | 73.2 | 0.79 | 0.18, 0.17 | 74.52 | 0.78 | 0.18, 0.16 | ||
KEMatch *Mention Para. Rep. | + | 77.41 | 0.86 | 0.17, 0.15 | 76.49 | 0.85 | 0.17, 0.16 | |
\(\odot\) | 80.13 | 0.87 | 0.16, 0.15 | 80.41 | 0.87 | 0.16, 0.16 | ||
\(-\) | 73.29 | 0.81 | 0.18, 0.18 | 72.48 | 0.82 | 0.19, 0.18 | ||
all | 74.19 | 0.81 | 0.18, 0.17 | 74.07 | 0.83 | 0.19, 0.18 | ||
KEMatch *Title + Mention Para. Rep. | + | 72.28 | 0.78 | 0.18, 0.16 | 71.78 | 0.77 | 0.19, 0.16 | |
KEMatch *Cxt. Query Ent Rep. | \(\odot\) | 81.2 | 0.87 | 0.15, 0.15 | 82.89 | 0.88 | 0.16, 0.15 | |
LTRMHSA | 84.13 | 0.89 | 0.14, 0.14 | 85.37 | 0.91 | 0.14, 0.14 |
“Produced by an Igloolik, Nunavut company, the film is titled The Journals of Knud Rassmusen, and co-directed by Zacharias Kunuk of Igloolik and Norman Cohn of Montreal. The company received critical acclaim for their first film, Atanarjuat, ... The film portrays the pressures on traditional Inuit culture....”4
“Firefighters have confirmed that the large James Bond sound stage at Pinewood Studios has been destroyed by fire... where filming for Casino Royale, the latest Bond movie, has been completed... Pinewood, which was created in 1935, was the filming ground for Dr No, the first ever James Bond movie in 1962.” 5
5.3 Comparison between Non-monotonic and Monotonic Path Generation
%Unf | %Recov | AVG * PW Sim | NGEO(E),NGEO(R) | |
---|---|---|---|---|
\(\pi ^*_{annealed}\) *(Non-Monotonic) | 0 | 84.13 | 0.89 | 0.14, 0.14 |
\(\pi ^*_{L\text{-}R}\) * (Monotonic) | 29.73 | 42.85 | 0.73 | 0.19, 0.16 |
“...Ledger starred in the 2005 movie Brokeback Mountain where he was nominated for the Academy Award and the Golden Globe Award for Best Actor. He also starred in the 2000 movie The Patriot with Mel Gibson ...” 6
6 Discussion
7 Conclusion
Acknowledgments
Footnotes
A Experiments On Synthetic Dataset
Synthetic Dataset | |
# context documents | 2,000 |
# entities mentioned | 19,173 |
# entity pairs | 5,000 |
AVG/MAX ground truth path length | 4/6 |
Knowledge Graph | |
# entities | 59,173 |
# relations | 651 |
Ground Truth Paths | |
AVG path length | 4 |
# unique path entities | 19,173 |
# unique path relations | 648 |
A.1 Coping with Large Dataset
Dataset | Synthetic | ||||
---|---|---|---|---|---|
Feature * Combination | %Recov | AVG * PW Sim | NGEO(E), NGEO(R) | ||
NCNMPG | 16.2 | 0.32 | 0.32, 0.25 | ||
NMCPGT | Random Context | 55.43 | 0.47 | 0.29, 0.23 | |
Context Window | 61.83 | 0.56 | 0.25, 0.22 | ||
KEMatch * AVG Ent Rep. | + | 63.46 | 0.68 | 0.25, 0.22 | |
\(\odot\) | 64.27 | 0.68 | 0.24, 0.2 | ||
\(-\) | 57.2 | 0.55 | 0.27, 0.23 | ||
all | 62.35 | 0.64 | 0.25, 0.22 | ||
KEMatch * Mention Para. Rep. | + | 63.59 | 0.65 | 0.25, 0.21 | |
\(\odot\) | 73.66 | 0.71 | 0.21, 0.18 | ||
\(-\) | 58.74 | 0.53 | 0.26, 0.21 | ||
all | 61.57 | 0.58 | 0.26, 0.21 | ||
KEMatch * Title + Mention Para. Rep. | + | 60.91 | 0.58 | 0.27, 0.21 | |
KEMatch * Cxt. Query Ent Rep. | \(\odot\) | 75.37 | 0.74 | 0.18, 0.17 | |
LTRMHSA | 75.92 | 0.76 | 0.17, 0.16 |
A.2 Inferring New Relation Edges
k | %Recov | AVG PW * Sim | NGEO(E),NGEO(R) | % \(E_{\text{CP}}^-\) | % \(E_{\text{CP}}^-\) * Recov |
---|---|---|---|---|---|
0 * (Full KG) | 84.13 | 0.89 | 0.14, 0.14 | - | - |
10 | 80.22 | 0.86 | 0.15, 0.14 | 5.37 | 65.92 |
30 | 65.23 | 0.71 | 0.19, 0,17 | 19.31 | 57.23 |
50 | 47.37 | 0.59 | 0.24, 0.23 | 38.45 | 40.85 |
References
Index Terms
- Non-monotonic Generation of Knowledge Paths for Context Understanding
Recommendations
A transformer framework for generating context-aware knowledge graph paths
AbstractContextual Path Generation (CPG) refers to the task of generating knowledge path(s) between a pair of entities mentioned in an input textual context to determine the semantic connection between them. Such knowledge paths, also called contextual ...
Evaluating the Impact of Knowledge Graph Context on Entity Disambiguation Models
CIKM '20: Proceedings of the 29th ACM International Conference on Information & Knowledge ManagementPretrained Transformer models have emerged as state-of-the-art approaches that learn contextual information from the text to improve the performance of several NLP tasks. These models, albeit powerful, still require specialized knowledge in specific ...
Leveraging Knowledge Context Information to Enhance Personalized Recommendation
Neural Information ProcessingAbstractKnowledge graphs (KGs) have proven to be effective to improve the performance of recommendation. However, with the tremendous increase of users and items, existing methods still face several challenging problems: (1) path-based methods rely ...
Comments
Information & Contributors
Information
Published In
Publisher
Association for Computing Machinery
New York, NY, United States
Publication History
Check for updates
Author Tags
Qualifiers
- Research-article
Funding Sources
- Lee Kong Chian Professorship and National Research Foundation, Singapore
- Strategic Capabilities Research Centres Funding Initiative
Contributors
Other Metrics
Bibliometrics & Citations
Bibliometrics
Article Metrics
- 0Total Citations
- 555Total Downloads
- Downloads (Last 12 months)555
- Downloads (Last 6 weeks)63
Other Metrics
Citations
View Options
Get Access
Login options
Check if you have access through your login credentials or your institution to get full access on this article.
Sign in