1 Introduction
2 Problem Definition
3 Estimation of Entity Typicality
4 Measurement of Temporal Comparability
One, new, two, like, people, | |
Year, many, women, time, company, | |
Work, city, water, make, way, | |
Use, world, business, school, life |
5 ILP Formulation for Detecting Comparables
6 Experiments
6.1 Datasets
Period | LOC | PRODUCT | NORP | WOA | GPE | PERSON | FACT | ORG |
---|---|---|---|---|---|---|---|---|
\(T_{A1}\)
| 427 | 87 | 2959 | 129 | 7810 | 33,127 | 328 | 23,775 |
\(T_{A2}\)
| 370 | 57 | 2203 | 103 | 4698 | 27,932 | 247 | 16,751 |
\(T_{B}\)
| 304 | 44 | 1573 | 91 | 4460 | 16,103 | 221 | 11,215 |
Period | LAW | EVENT | TOTAL | |||||
---|---|---|---|---|---|---|---|---|
\(T_{A1}\)
| 18 | 212 | 68,872 | |||||
\(T_{A2}\)
| 10 | 149 | 52,520 | |||||
\(T_{B}\)
| 11 | 129 | 34,151 |
6.2 Test Sets
6.3 Evaluation Criteria
6.3.1 Criteria for Quantitative Evaluation
6.3.2 Criteria for Qualitative Evaluation
6.4 Baselines
6.5 Experiment Settings
6.6 Evaluation Results
6.6.1 Results of Quantitative Evaluation
Method | Embedding-S+Non-Tran | Embedding-J | Embedding-S+OT | ||||||
---|---|---|---|---|---|---|---|---|---|
Prec. | Recall | \(F_{1}\)-score | Prec. | Recall | \(F_{1}\)-score | Prec. | Recall | \(F_{1}\)-score | |
K-means |
0.027
|
0.030
|
0.028
| 0.081 | 0.089 | 0.085 | 0.186 | 0.195 | 0.190 |
DBSCAN | 0.000 | 0.000 | 0.000 | 0.000 | 0.000 | 0.000 | 0.105 | 0.106 | 0.105 |
AP | 0.016 | 0.018 | 0.017 | 0.049 | 0.054 | 0.051 | 0.154 | 0.160 | 0.156 |
MRRW | 0.000 | 0.000 | 0.000 | 0.027 | 0.030 | 0.028 | 0.132 | 0.136 | 0.133 |
I-ILP | 0.016 | 0.018 | 0.017 | 0.049 | 0.054 | 0.051 | 0.165 | 0.171 | 0.167 |
J-ILP | 0.000 | 0.000 | 0.000 |
0.124
|
0.137
|
0.130
|
0.273
|
0.290
|
0.281
|
Method | Embedding-S+Non-Tran | Embedding-J | Embedding-S+OT | ||||||
---|---|---|---|---|---|---|---|---|---|
Prec. | Recall | \(F_{1}\)-score | Prec. | Recall | \(F_{1}\)-score | Prec. | Recall | \(F_{1}\)-score | |
K-means | 0.005 | 0.009 | 0.006 | 0.038 | 0.064 | 0.048 | 0.146 | 0.248 | 0.184 |
DBSCAN | 0.000 | 0.000 | 0.000 | 0.027 | 0.046 | 0.034 | 0.092 | 0.156 | 0.116 |
AP | 0.000 | 0.000 | 0.000 | 0.016 | 0.028 | 0.020 | 0.141 | 0.239 | 0.177 |
MRRW | 0.000 | 0.000 | 0.000 | 0.016 | 0.028 | 0.020 | 0.157 | 0.266 | 0.197 |
I-ILP |
0.011
|
0.018
|
0.017
| 0.043 | 0.073 | 0.054 | 0.151 | 0.257 | 0.190 |
J-ILP | 0.000 | 0.000 | 0.000 |
0.059
|
0.101
|
0.075
|
0.178
|
0.303
|
0.224
|
Entity pair | K-means | DBSCAN | AP | MRRW | I-ILP | J-ILP |
---|---|---|---|---|---|---|
(iraq, syria) | (0,0) | (1,1)* | (1,0) | (1,0) | (1,1)* | (1,1) |
(president_reagan, george_bush) | (1,1)* | (0,1) | (0,1) | (0,1) | (1,0) | (1,1) |
(american_express, credit_card) | (1,0) | (0,0) | (0,0) | (0,0) | (1,1) | (1,1) |
(macintosh, pc) | (1,1) | (1,0) | (0,0) | (0,0) | (1,0) | (1,0) |
(salomon, morgan_stanley) | (0,1) | (0,0) | (1,0) | (1,0) | (0,1) | (1,1) |
(national_basketball, world_series) | (1,1) | (0,0) | (0,1) | (0,0) | (0,1) | (0,1) |
(european_community, china) | (0,1) | (0,0) | (0,1) | (1,0) | (0,0) | (0,1) |
(pan_am, american_airlines) | (1,1)* | (1,0) | (1,1)* | (0,0) | (1,1) | (1,0) |
(mario_cuomo, george_pataki) | (0,1) | (1,0) | (0,1) | (0,0) | (1,1)* | (1,1) |
(bonn, berlin) | (0,0) | (0,0) | (1,0) | (1,0) | (1,1) | (1,1) |
(sampras, federer) | (0,0) | (1,1) | (0,0) | (0,0) | (0,1) | (0,0) |
(saddam, al_qaeda) | (1,1) | (1,0) | (1,0) | (0,1) | (0,0) | (1,0) |
6.6.2 Results of Qualitative Evaluation
6.7 Additional Observations
6.7.1 Additional Metrics
Model | Embedding-S+Non-Tran | Embedding-J | Embedding-S+OT | ||||||
---|---|---|---|---|---|---|---|---|---|
Typ | Comp | Product | Typ | Comp | Product | Typ | Comp | Product | |
K-means | 50.505 | 2.361 | 119.242 | 57.470 | 2.680 | 154.020 | 62.052 | 3.263 | 202.476 |
DBSCAN | 49.504 | 2.320 | 114.849 | 56.862 | 3.495 | 198.733 | 61.092 | 3.786 | 231.294 |
AP | 50.573 | 2.333 | 117.987 | 57.635 | 3.105 | 178.957 | 61.743 | 3.562 | 219.929 |
MRRW | 46.043 | 2.069 | 95.263 | 48.136 | 2.348 | 113.023 | 56.588 | 3.534 | 199.982 |
I-ILP |
51.037
| 2.167 | 110.597 |
58.592
| 2.680 | 157.027 |
62.799
| 3.339 | 209.686 |
J-ILP | 49.467 |
3.186
|
157.602
| 56.207 |
3.712
|
208.640
| 60.752 |
3.903
|
237.115
|
6.7.2 Effects of the Number of CFTs
6.7.3 Effects of Trade-Off Parameter
6.7.4 Sensitivity to Kernel Choice
Kernel function | Quatic | Triweight | Epanechnikov | Cosine |
---|---|---|---|---|
Difference rate | 15.9 | 10.3 | 5.5 | 15.9 |
7 Related Work
8 Discussions
-
In the problem setting, the compared entity sets are extracted from two different time periods. However, our proposed J-ILP model also works for mapping two entity sets from the same time period (e.g., comparing European politicians with contemporary Asian politicians). Note that in this case, we do not need to solve the across-time context alignment problem. Entity vectors from different sets can be compared directly based on their cosine similarity.
-
On the other hand, when mapping entity sets across time, the same entity which appears in both time periods may be discovered and paired together in the result, in case that such entity is associated with a stable diachronic meaning (e.g., New York, dollar etc.). However, entities with changed roles will less likely be included in the result, since the temporal comparability between their meaning at different times is low. In this task, we focus more on generating similar entity pairs that are beneficial for understanding the connection between two time periods.
-
We show a few examples of mapped entity pairs in Table 5. For instance, (president_reagan, george_bush) and (mario_cuomo, george_pataki) represent the corresponding US President and the governor of New York in different periods, respectively. As another two examples, (bonn, berlin) represents the pair of German capital before and after the reunification of Western and Eastern German. (salomon, morgan_stanley) describes the pair of temporal corresponding large investment banks. We can see these detected pairs are clear and conveying comparative historical knowledge.
-
Our framework relies on the orthogonal transformation for computing across-time entity comparability and a ILP framework for generating exemplar entity pairs. However, the transfer of deep learning framework to our task may bring new insights. For example, we may use an approach similar to the one proposed in [31] for discovering across-time analogy relationships.
-
Finally, there are some connections between our research problems with the knowledge graph embedding task (KGE). A knowledge graph is a multi-relational graph composed of entities (nodes) and relations (different types of edges), and the goal of knowledge graph embedding is to embed components of a knowledge graph including entities and relations into continuous vector spaces, so as to simplify the manipulation while preserving the inherent structure of the knowledge graph [41]. In our task, we also focus on the triple of the form (entity from period\(T_{A}\), temporal analog, entity from period\(T_{B}\)), where temporal analog denotes the relation between entities from \(T_{A}\) and \(T_{B}\). However, the difference lies in that we focus on the construction of such triples from news archives, while KGE aims to learn effective representations based on these triples.