Improving Time Sensitivity for Question Answering over Temporal Knowledge Graphs

Chao ShangGuangtao WangPeng QiJing Huang

article2022ACL55 citations

Proposes a time-sensitive question answering framework that infers implicit timestamps and injects temporal ordering into knowledge graph embeddings, achieving a 32% absolute error reduction on multi-step temporal reasoning queries.

Listen

Modern decision-making systems increasingly rely on automated question answering over knowledge graphs to retrieve time-sensitive facts, such as leadership tenures, event sequences, and operational histories. However, existing temporal question answering systems face severe accuracy bottlenecks when processing complex natural language queries. These systems routinely struggle because questions frequently omit explicit dates (requiring unstated timeframes to be inferred), standard language models fail to register critical temporal relational prepositions (such as distinguishing "before" from "after"), and underlying mathematical graph representations treat timestamps as isolated symbols without recognizing chronological order or duration.

The article demonstrates and evaluates a time-sensitive question answering framework called TSQA, designed to significantly improve automated multi-step temporal reasoning across structured knowledge graphs.

The authors constructed a multi-stage framework that directly resolves earlier architectural gaps. The approach integrates an explicit time estimation module to infer missing reference dates, incorporates auxiliary time-order training constraints into graph embeddings to encode chronological progression, and introduces contrastive learning objectives that force the model to distinguish between opposing temporal phrasing. Additionally, the system extracts targeted subgraphs around mentioned entities to dramatically restrict the candidate search space during training and inference. The framework was evaluated on the large-scale benchmark CRONQUESTIONS, which encompasses 125,000 entities, 328,000 temporal facts, and 410,000 natural language questions.

The findings establish substantial performance advantages over existing state-of-the-art baselines. Overall, TSQA raised top-1 answer accuracy from 64.7% to 83.1% across the complete benchmark. On complex reasoning questions requiring multi-step fact integration, TSQA achieved a 32% absolute error reduction, increasing top-1 accuracy from 39.2% to 71.3%—an 82% relative improvement. Fine-grained evaluations revealed massive gains on challenging subtypes, boosting accuracy on "first/last" sequence questions by 94% and "before/after" queries by 75%, while maintaining near-perfect accuracy (98.7%) on simple direct queries. Ablation analysis confirmed that explicit time estimation was the single most impactful component, contributing an immediate 14.5% boost in top-1 accuracy on complex questions, followed by subgraph extraction, which improved accuracy by 7.8% while shrinking the average graph search space down to roughly 3% of the total dataset.

These results indicate that automated question answering over temporal databases cannot rely solely on standard text encoders and isolated graph embeddings; explicit chronological ordering and structured time estimation are essential. For organizations managing enterprise knowledge bases, intelligence archives, or compliance records, adopting this architecture significantly reduces the operational risk of erroneous time-based answers and false positives. It also delivers computational efficiencies by pruning the required graph search space prior to scoring.

Organizations developing or deploying knowledge graph search engines should transition from static graph embeddings to time-aware encoders and incorporate intermediate time-estimation steps in their query processing pipelines. Implementation teams should also adopt neighboring graph pruning to lower cloud compute overhead and latency in real-time inference environments.

Confidence in these findings is high for structured knowledge bases with discrete annual timestamps, given the rigorous benchmarking across hundreds of thousands of queries. However, decision-makers should note that the evaluation relied on data discretized by year; real-world applications involving fine-grained temporal data (such as exact dates, minutes, or continuously streaming intervals) will require further validation before full-scale deployment.

Amazon JD AI Research complex questions natural language questions reasoning temporal knowledge graph question answering temporal knowledge graphs temporal reasoning

arXiv: 2203.00255

Paper: A Survey on Knowledge Graphs: Representation, Acquisition, and Applications, Shaoxiong Ji et al. (2020). This survey supplies the essential taxonomy of knowledge-graph representations, temporal modeling, embeddings, and downstream question-answering applications that TSQA specializes for time-sensitive reasoning.
Paper: A Review of Relational Machine Learning for Knowledge Graphs, Maximilian Nickel et al. (2015). Its review of relational learning, latent embeddings, and graph-path reasoning provides the foundational modeling vocabulary for TSQA’s time-aware graph representations and subgraph search.
Paper: Knowledge Graph Embedding via Dynamic Mapping Matrix, Guoliang Ji et al. (2015). TransD’s dynamic entity-relation projections clarify the embedding machinery that TSQA extends by encoding chronological order rather than treating timestamps as isolated symbols.
Paper: Semantic Parsing on Freebase from Question-Answer Pairs, Jonathan Berant et al. (2013). SEMPRE establishes the semantic-parsing approach to mapping natural-language questions into executable knowledge-base operations that TSQA augments with temporal estimation and ordering.
Paper: Relevance-Based Language Models, Victor Lavrenko et al. (2001). This paper provides an early principled treatment of timestamps in retrieval models, preparing the idea that temporal information should directly influence relevance rather than serve as an isolated filter.

Paper: Plan-on-Graph: Self-Correcting Adaptive Planning of Large Language Model on Knowledge Graphs, Liyi Chen et al. (2024). Plan-on-Graph extends structured question answering beyond TSQA’s temporal subgraph reasoning by adding adaptive exploration, dynamic memory, backtracking, and self-correction over knowledge graphs.
Paper: Unifying Large Language Models and Knowledge Graphs: A Roadmap, Shirui Pan et al. (2023). This roadmap generalizes the source’s time-aware graph-question-answering architecture into a broader program for integrating language models and knowledge graphs through retrieval, prompting, and bidirectional reasoning.

Cover for Improving Time Sensitivity for Question Answering over Temporal Knowledge Graphs

Abstract

Question answering over temporal knowledge graphs (KGs) efficiently uses facts contained in a temporal KG, which records entity relations and when they occur in time, to answer natural language questions (e.g., “Who was the president of the US before Obama?”). These questions often involve three time-related challenges that previous work fail to adequately address: 1) questions often do not specify exact timestamps of interest (e.g., “Obama” instead of 2000); 2) subtle lexical differences in time relations (e.g., “before” vs “after”); 3) off-the-shelf temporal KG embeddings that previous work builds on ignore the temporal order of timestamps, which is crucial for answering temporal-order related questions. In this paper, we propose a time-sensitive question answering (TSQA) framework to tackle these problems. TSQA features a timestamp estimation module to infer the unwritten timestamp from the question. We also employ a time-sensitive KG encoder to inject ordering information into the temporal KG embeddings that TSQA is based on. With the help of techniques to reduce the search space for potential answers, TSQA significantly outperforms the previous state of the art on a new benchmark for question answering over temporal KGs, especially achieving a 32% (absolute) error reduction on complex questions that require multiple steps of reasoning over facts in the temporal KG.

Topics

complex questions natural language questions reasoning timestamping estimation method temporal correlations bipartite graphs knowledge graph embedding

Keywords

temporal knowledge graph question answering temporal knowledge graphs temporal reasoning time-sensitive question answering temporal contrastive learning temporal knowledge graph embeddings timestamp embeddings timestamp estimation

1 Introduction
2 Related Work
3 Method
3.1 Problem Definition and Framework
3.2 Time-aware KG Encoder
3.3 Time-Sensitive TKG-QA
3.3.1 Question Decomposition and Encoder
3.3.2 Entity Neighbor Graph Extraction
3.3.3 Time-Sensitive Question Answering
3.3.4 Temporal Contrastive Learning
4 Experiments
4.1 Experimental Setup
4.2 Main Results
4.3 Ablation Study
5 Conclusion
Acknowledgements
References

Knowls

Knowl 1 — Time-Sensitive Question Answering (TSQA) Framework for Temporal Knowledge Graphs
model/method
Temporal Knowledge Graph Question Answering (TKGQA) aims to find an answer entity $v \in V$ or timestamp $t \in T$ from a temporal knowledge graph $G = (V, E, R, T)$ given a natural language question $Q$ containing implicit temporal expressions. In $G$ , quadruples $(s, r, [t_s, t_e], o)$ specify that relation $r \in R$ holds between subject entity $s \in V$ and object entity $o \in V$ over the time interval $[t_s, t_e]$ with $t_s, t_e \in T$ .

The TSQA framework addresses multi-hop and temporal-ordering reasoning via two primary components:
1. Time-Aware TKG Encoder: Learns complex-valued representations of entities, relations, and timestamps by combining TCompLEx score modeling with an auxiliary pairwise time-order classification objective that enforces chronological order and relative temporal distance in the embedding space.
2. Time-Sensitive Question Answering Module:
  - Decomposes the input question $Q$ into an entity-independent temporal expression $\hat{Q}$ by substituting recognized KG entities with placeholder tokens ([subject], [object]), then encodes $\hat{Q}$ with BERT to generate complex relation ( $q_r$ ) and time ( $q_t$ ) question representations.
  - Extracts an $m$ -hop candidate search subgraph $G_q$ surrounding question entities to restrict candidate entities $E_q$ and candidate timestamps $T_q$ .
  - Performs time estimation to infer a latent question timestamp embedding $t_q$ , which is directly fed into the entity prediction module to score candidate entities alongside candidate timestamps.
  - Applies temporal contrastive learning using antonym question substitutions to heighten model sensitivity to directional temporal words (such as "before" versus "after").
Knowl 2 — Time-Aware Temporal KG Embedding with Auxiliary Time-Order Learning
model/method
To preserve chronological order and distance information in embedding space, TSQA augments TCompLEx embedding learning with an auxiliary time-order classification objective.

First, interval quadruples $(s, r, [t_s, t_e], o)$ are expanded into individual timestamp quadruples:

$(s, r, [t_s, t_e], o) = \{ (s, r, t, o) \mid t_s \le t \le t_e \}$

Let $e_s, e_r, e_o, e_t \in \mathbb{C}^d$ denote the complex embeddings of subject $s$ , relation $r$ , object $o$ , and timestamp $t \in T$ . TCompLEx computes the fact score as:

$S(s, r, o, t) = \text{Re}(\langle e_s, e_r, e_o, e_t \rangle)$

where $\text{Re}(\cdot)$ denotes the real part of a complex vector and $\langle \cdot \rangle$ denotes the multilinear product. The base loss with $K$ negative samples is:

$L_{TC} = -\log\left(\phi(\gamma - S(s, r, o, t))\right) - \frac{1}{K} \sum_{i=1}^K \log\left(\phi(S(s'_i, r, o'_i, t'_i) - \gamma)\right)$

where $\gamma$ is a fixed margin and $\phi$ is the sigmoid function.

To encode order, timestamps in $T$ are sorted in ascending order $(t_1, t_2, \dots, t_{|T|})$ such that $t_i < t_j$ for $1 \le i < j \le |T|$ . Let $\mathbf{t}_i = [\text{Re}(e_{t_i}), \text{Im}(e_{t_i})] \in \mathbb{R}^{2d}$ denote the concatenation of real and imaginary parts of timestamp embedding $e_{t_i}$ . Timestamp embeddings are initialized with sinusoidal positional representations for $0 \le k \le d - 1$ :

$\mathbf{t}_i[2k] = \sin\left( \frac{i}{10000^{2k/(2d)}} \right), \quad \mathbf{t}_i[2k+1] = \cos\left( \frac{i}{10000^{2k/(2d)}} \right)$

For any timestamp pair $(t_i, t_j)$ , the predicted probability of $t_i$ preceding $t_j$ is:

$p_t(i, j) = \text{sigmoid}((\mathbf{t}_i - \mathbf{t}_j)^T W_t)$

where $W_t \in \mathbb{R}^{2d}$ is a parameter vector. The time-order loss is:

$L_{TO} = -\delta(i, j) \log(p_t(i, j)) - (1 - \delta(i, j)) \log(1 - p_t(i, j))$

where $\delta(i, j) = 1$ if $t_i < t_j$ , and $\delta(i, j) = 0$ otherwise. The TKG encoder is trained on a weighted combination of $L_{TC}$ and $L_{TO}$ .
Knowl 3 — Question Decomposition and Complex Vector Formulation in TSQA
model/method
TSQA separates entity mentions from temporal-relational text in questions and maps language model token representations into complex vector spaces compatible with TKG embeddings:
1. Entity Masking: Given a question $Q$ containing entities $\{\text{Ent}_1, \dots, \text{Ent}_k\} \subseteq V$ , each entity is replaced in order with special tokens [subject] and [object], creating an entity-independent temporal expression $\hat{Q}$ (e.g., "When did Obama hold the position of President of USA?" becomes "When did [subject] hold the position of [object]?").
2. Text Encoding: The sequence [CLS] + $\hat{Q}$ is passed through BERT, yielding the [CLS] embedding $e_q \in \mathbb{R}^{d_{\text{bert}}}$ .
3. Projection to Relation and Time Spaces: Two real projection heads map $e_q$ to relation and time vector spaces:
$q_r = W_q^r(\delta(W e_q)) \in \mathbb{R}^{2d}$

$q_t = W_q^t(\delta(W e_q)) \in \mathbb{R}^{2d}$

where $W \in \mathbb{R}^{2d \times d_{\text{bert}}}$ , $W_q^r, W_q^t \in \mathbb{R}^{2d \times 2d}$ , and $\delta$ is an activation function. 4. Complex Formulation: Real vectors $q_r, q_t$ are converted to complex vectors $q_r, q_t \in \mathbb{C}^d$ by partitioning real and imaginary segments:

$q_r = q_r[0:d] + \sqrt{-1} \cdot q_r[d:2d]$

$q_t = q_t[0:d] + \sqrt{-1} \cdot q_t[d:2d]$
Knowl 4 — Candidate Search Space Pruning via Entity Neighbor Subgraph Extraction
model/method
To reduce candidate space size and eliminate spurious distractor candidates during inference, TSQA constrains candidate answers to a localized neighborhood subgraph.

For a question $Q$ with identified entities $\{\text{Ent}_1, \dots, \text{Ent}_k\}$ , an $m$ -hop neighboring subgraph $G_i$ is extracted from the temporal KG $G = (V, E, R, T)$ for each entity $\text{Ent}_i$ . The combined question search graph is:

$G_q = \bigcup_{i=1}^k G_i$

The candidate entity set $E_q \subseteq V$ and candidate timestamp set $T_q \subseteq T$ are defined as the entities and timestamps present within $G_q$ .
- Training stage: The hop number $m$ is set to the minimum value that includes the ground-truth answer entity/timestamp in $G_q$ .
- Testing stage: $m$ is set to the maximum hop number encountered during training.
On the CRONQUESTIONS benchmark, this reduces the average candidate search space $|G_q|/|G|$ to approximately 3% of the full knowledge graph.
Knowl 5 — Joint Time Estimation and Answer Prediction Scoring Formulation
equation
In TSQA, time estimation produces a latent reference time representation that directly conditions entity prediction:
1. Time Estimation: Given subject entity embedding $e_s \in \mathbb{C}^d$ , object entity embedding $e_o \in \mathbb{C}^d$ , and question time embedding $q_t \in \mathbb{C}^d$ , the estimated question time representation $t_q \in \mathbb{C}^d$ is computed by:
$t_q = F_T(e_s, q_t, e_o) = W_q^t \left( [\text{Re}(\langle e_s, q_t, e_o \rangle), \text{Im}(\langle e_s, q_t, e_o \rangle)] \right)$

where $W_q^t \in \mathbb{R}^{2d \times 2d}$ , $[\cdot]$ denotes vector concatenation, and $\text{Re}(\cdot), \text{Im}(\cdot)$ denote real and imaginary components. (If the question contains an explicit timestamp, $t_q$ is a linear combination of this estimated vector and the KG timestamp embedding.) Candidate timestamps $t \in T_q$ receive score:

$S_t = \text{Re}(\langle t_q, t \rangle)$
1. Entity Prediction: The inferred time vector $t_q$ is passed to the entity function $F_E$ alongside question relation embedding $q_r \in \mathbb{C}^d$ and subject embedding $e_s$ :
$e_q = F_E(e_s, q_r, t_q) = \langle e_s, q_r, t_q \rangle \in \mathbb{C}^d$

Candidate entities $e \in E_q$ receive score:

$S_e = \text{Re}(\langle e_q, e \rangle)$
1. Answer Loss: Over total candidates $C = |E_q| + |T_q|$ with candidate scores $S_{a,i} \in \{S_e, S_t\}$ , the candidate probability and cross-entropy loss are:
$P_{a,i} = \frac{\exp(S_{a,i})}{\sum_{j=1}^C \exp(S_{a,j})}, \quad L_{\text{answer}} = -\sum_{i=1}^C y_i \log(P_{a,i})$

where $y_i = 1$ for the ground-truth answer and $0$ otherwise.
Knowl 6 — Temporal Contrastive Learning with Dual Time-Order and Answer Guidance
model/method
To enhance model sensitivity to temporal relation words in questions, TSQA generates contrastive question pairs using antonym substitutions and optimizes two auxiliary contrastive objectives:
1. Contrastive Question Construction: A temporal word in question $Q$ is replaced with its antonym from the predefined dictionary:
$\mathcal{D}_{\text{contr}} = \{(\text{first}, \text{last}), (\text{before}, \text{after}), (\text{before}, \text{during}), (\text{during}, \text{after}), (\text{before}, \text{when}), (\text{when}, \text{after})\}$

yielding contrastive question $\bar{Q}$ .
1. Contrastive Time Order Learning: Time embeddings $t_q$ and $t_{q_c}$ are generated for $Q$ and $\bar{Q}$ . With binary order label $y_o = 0$ if word 1 was replaced by word 2 and $y_o = 1$ otherwise, the model predicts the order probability $p_o$ and minimizes $L_{\text{order}}$ :
$p_o = \text{sigmoid}((t_q - t_{q_c})^T W_o), \quad L_{\text{order}} = -y_o \log(p_o) - (1 - y_o) \log(1 - p_o)$

where $W_o \in \mathbb{R}^{2d}$ is a learnable vector.
1. Answer-Guided Contrastive Learning: Candidate answer score vectors $S, \bar{S} \in \mathbb{R}^C$ for $Q$ and $\bar{Q}$ are stacked into $S_q = [S; \bar{S}] \in \mathbb{R}^{2 \times C}$ . Applying column-wise softmax gives normalized probability matrix $P_q = \text{softmax}(S_q) \in \mathbb{R}^{2 \times C}$ . Because the true answer for $Q$ is invalid for $\bar{Q}$ , the contrastive loss penalizes high scores for $Q$ 's answer under $\bar{Q}$ :
$L_{\text{contrast}} = -\frac{1}{C} \sum_{i=1}^C y_i \log(P_q[0, i])$
1. Joint Training Loss:
$\text{Loss} = L_{\text{answer}} + \lambda_o L_{\text{order}} + \lambda_c L_{\text{contrast}}$

where $\lambda_o, \lambda_c > 0$ are balancing hyperparameters.

Knowl 7 — Performance of TSQA on the CRONQUESTIONS Benchmark

data/table

The performance of TSQA compared against baseline temporal and non-temporal KGQA models on the CRONQUESTIONS test set (125k entities, 328k quadruples, 30k test questions) evaluated by Hits@1 and Hits@10:

Model	Hits@1					Hits@10
	Overall	Question Type		Answer Type		Overall	Question Type		Answer Type
	Overall	Complex	Simple	Entity	Time	Overall	Complex	Simple	Entity	Time
EmbedKGQA	0.288	0.286	0.290	0.411	0.057	0.672	0.632	0.725	0.850	0.341
T-EaE-add	0.278	0.257	0.306	0.313	0.213	0.663	0.614	0.729	0.662	0.665
T-EaE-replace	0.288	0.257	0.329	0.318	0.231	0.678	0.623	0.753	0.668	0.698
CronKGQA	0.647	0.392	0.987	0.699	0.549	0.884	0.802	0.992	0.898	0.857
TSQA	0.831	0.713	0.987	0.829	0.836	0.980	0.968	0.997	0.981	0.978

TSQA improves overall Hits@1 from 0.647 to 0.831 and Hits@10 from 0.884 to 0.980 relative to CronKGQA. On complex temporal reasoning questions requiring multiple relational/temporal steps, TSQA achieves an 82% relative improvement in Hits@1 (0.713 vs. 0.392, corresponding to a 32% absolute error reduction), while maintaining top accuracy on simple single-fact questions (0.987 Hits@1).

Knowl 8 — Subtype Question Accuracy Breakdown on CRONQUESTIONS

data/table

Hits@1 performance across fine-grained question subtypes on the CRONQUESTIONS test set:

Question Type	Before/After	First/Last	Time Join	Simple Entity	Simple Time
EmbedKGQA	0.199	0.324	0.223	0.421	0.087
T-EaE-add	0.256	0.285	0.175	0.296	0.321
T-EaE-replace	0.256	0.288	0.168	0.318	0.346
CronKGQA	0.288	0.371	0.511	0.988	0.985
TSQA	0.504	0.721	0.799	0.988	0.987

TSQA substantially outperforms previous models on complex question subtypes:

Before/After: Hits@1 increases from 0.288 to 0.504 (75.0% relative improvement).
First/Last: Hits@1 increases from 0.371 to 0.721 (94.3% relative improvement).
Time Join: Hits@1 increases from 0.511 to 0.799 (56.4% relative improvement). Accuracy on simple entity (0.988) and simple time (0.987) questions matches the performance of CronKGQA.

Knowl 9 — Ablation Analysis of TSQA Modules

empirical result

Ablation of TSQA components on CRONQUESTIONS test set, removing Temporal Contrastive Learning (TC), Time-Aware TKG Embeddings (TKE), Neighboring Graph Extraction (NG), and Time Estimation (TE) sequentially:

Model Variant	Overall	Question Type		Answer Type
Model Variant	Overall	Complex	Simple	Entity	Time
TSQA	0.831	0.713	0.987	0.829	0.836
- TC	0.821	0.696	0.984	0.820	0.822
- TC - TKE	0.816	0.688	0.985	0.816	0.818
- TC - TKE - NG	0.757	0.583	0.986	0.797	0.687
- TC - TKE - NG - TE	0.661	0.412	0.989	0.719	0.556

Component contributions:

Time Estimation (TE): Removing TE incurs the largest performance drop, reducing complex Hits@1 from 0.583 to 0.412 (-17.1% absolute) and overall Hits@1 from 0.757 to 0.661 (-9.6% absolute), confirming that intermediate time representations provide an essential anchor for answer entity retrieval.
Neighboring Graph Extraction (NG): Removing candidate pruning drops complex Hits@1 from 0.688 to 0.583 (-10.5% absolute) and overall Hits@1 by 5.9% absolute, indicating that candidate search reduction eliminates distractor errors.
TKE & Contrastive Learning (TC): Adding TKE and TC progressively improves complex question Hits@1 from 0.688 to 0.696 and 0.713, respectively, verifying that explicit order modeling in both KG embeddings and question representations improves sensitivity to temporal sequence constraints.

Coverage note — None was omitted; all key architectural components, mathematical formulations, loss functions, benchmark evaluations, subtype breakdowns, and ablation analyses were included.

References

1.Antoine Bordes, Nicolas Usunier, Alberto Garcia-Duran, Jason Weston, and Oksana Yakhnenko. 2013. Translating embeddings for modeling multi-relational data. Advances in neural information processing systems, 26.
2.Shib Sankar Dasgupta, Swayambhu Nath Ray, and Partha Talukdar. 2018. Hyte: Hyperplane-based temporally aware knowledge graph embedding. In Proceedings of the 2018 conference on empirical methods in natural language processing, pages 2001–2011.
3.Tim Dettmers, Pasquale Minervini, Pontus Stenetorp, and Sebastian Riedel. 2018. Convolutional 2d knowledge graph embeddings. In Thirty-second AAAI conference on artificial intelligence.
4.Bhuwan Dhingra, Jeremy R Cole, Julian Martin Eisenschlos, Daniel Gillick, Jacob Eisenstein, and William W Cohen. 2021. Time-aware language models as temporal knowledge bases. arXiv preprint arXiv:2106.15110.
5.Thibault Févry, Livio Baldini Soares, Nicholas FitzGerald, Eunsol Choi, and Tom Kwiatkowski. 2020. Entities as experts: Sparse memory access with entity supervision. arXiv preprint arXiv:2004.07202.
6.Alberto García-Durán, Sebastijan Dumančić, and Mathias Niepert. 2018. Learning sequence encoders for temporal knowledge graph completion. arXiv preprint arXiv:1809.03202.
7.Rishab Goel, Seyed Mehran Kazemi, Marcus Brubaker, and Pascal Poupart. 2020. Diachronic embedding for temporal knowledge graph completion. In Proceedings of the AAAI Conference on Artificial Intelligence.
8.Rujun Han, Xiang Ren, and Nanyun Peng. 2021. Econet: Effective continual pretraining of language models for event temporal reasoning. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 5367–5380.
9.Xiao Huang, Jingyuan Zhang, Dingcheng Li, and Ping Li. 2019. Knowledge graph embedding based question answering. In Proceedings of the Twelfth ACM International Conference on Web Search and Data Mining, pages 105–113.
10.Shaoxiong Ji, Shirui Pan, Erik Cambria, Pekka Marttinen, and S Yu Philip. 2021. A survey on knowledge graphs: Representation, acquisition, and applications. IEEE Transactions on Neural Networks and Learning Systems.
11.Zhen Jia, Abdalghani Abujabal, Rishiraj Saha Roy, Jannik Strötgen, and Gerhard Weikum. 2018a. Tempquestions: A benchmark for temporal question answering. In Companion Proceedings of the The Web Conference 2018, pages 1057–1062.
12.Zhen Jia, Abdalghani Abujabal, Rishiraj Saha Roy, Jannik Strötgen, and Gerhard Weikum. 2018b. Tequila: Temporal question answering over knowledge bases. In Proceedings of the 27th ACM International Conference on Information and Knowledge Management, pages 1807–1810.
13.Zhen Jia, Soumajit Pramanik, Rishiraj Saha Roy, and Gerhard Weikum. 2021. Complex temporal question answering on knowledge graphs. In Proceedings of the 30th ACM International Conference on Information & Knowledge Management, pages 792–802.
14.Tingsong Jiang, Tianyu Liu, Tao Ge, Lei Sha, Baobao Chang, Sujian Li, and Zhifang Sui. 2016. Towards time-aware knowledge graph completion. In Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers, pages 1715–1724.
15.Woojeong Jin, Rahul Khanna, Suji Kim, Dong-Ho Lee, Fred Morstatter, Aram Galstyan, and Xiang Ren. 2021. Forecastqa: A question answering challenge for event forecasting with temporal text data. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 4636–4650.
16.Timothée Lacroix, Guillaume Obozinski, and Nicolas Usunier. 2020. Tensor decompositions for temporal knowledge base completion. arXiv preprint arXiv:2004.04926.
17.Qiang Ning, Hao Wu, Rujun Han, Nanyun Peng, Matt Gardner, and Dan Roth. 2020. Torque: A reading comprehension dataset of temporal ordering questions. arXiv preprint arXiv:2005.00242.
18.Apoorv Saxena, Soumen Chakrabarti, and Partha Talukdar. 2021. Question answering over temporal knowledge graphs. arXiv preprint arXiv:2106.01515.
19.Apoorv Saxena, Aditay Tripathi, and Partha Talukdar. 2020. Improving multi-hop question answering over knowledge graphs using knowledge base embeddings. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 4498–4507.
20.Michael Schlichtkrull, Thomas N Kipf, Peter Bloem, Rianne Van Den Berg, Ivan Titov, and Max Welling. 2018. Modeling relational data with graph convolutional networks. In European semantic web conference, pages 593–607. Springer.
21.Chao Shang, Peng Qi, Guangtao Wang, Jing Huang, Youzheng Wu, and Bowen Zhou. 2021. Open temporal relation extraction for question answering. In 3rd Conference on Automated Knowledge Base Construction.
22.Chao Shang, Yun Tang, Jing Huang, Jinbo Bi, Xiaodong He, and Bowen Zhou. 2019. End-to-end structure-aware convolutional networks for knowledge base completion. In Proceedings of the AAAI Conference on Artificial Intelligence.
23.Zhiqing Sun, Zhi-Hong Deng, Jian-Yun Nie, and Jian Tang. 2019. Rotate: Knowledge graph embedding by relational rotation in complex space. arXiv preprint arXiv:1902.10197.
24.Yun Tang, Jing Huang, Guangtao Wang, Xiaodong He, and Bowen Zhou. 2019. Orthogonal relation transforms with graph context modeling for knowledge graph embedding. arXiv preprint arXiv:1911.04910.
25.Théo Trouillon, Johannes Welbl, Sebastian Riedel, Éric Gaussier, and Guillaume Bouchard. 2016. Complex embeddings for simple link prediction. In International conference on machine learning, pages 2071–2080. PMLR.
26.Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. In Advances in neural information processing systems, pages 5998–6008.
27.Jiapeng Wu, Meng Cao, Jackie Chi Kit Cheung, and William L Hamilton. 2020. Temp: temporal message passing for temporal knowledge graph completion. arXiv preprint arXiv:2010.03526.
28.Bishan Yang, Wen-tau Yih, Xiaodong He, Jianfeng Gao, and Li Deng. 2014. Embedding entities and relations for learning and inference in knowledge bases. arXiv preprint arXiv:1412.6575.

Citation

MLA

Shang, C., et al. “Improving Time Sensitivity for Question Answering over Temporal Knowledge Graphs”. Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 2022, pp. 8017–26, https://doi.org/10.18653/v1/2022.acl-long.552.

APA

Shang, C., Wang, G., Qi, P., & Huang, J. (2022). Improving Time Sensitivity for Question Answering over Temporal Knowledge Graphs. Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 8017–8026. https://doi.org/10.18653/v1/2022.acl-long.552

Chicago

Shang, C., G. Wang, P. Qi, and J. Huang. 2022. “Improving Time Sensitivity for Question Answering over Temporal Knowledge Graphs”. Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 8017–26. https://doi.org/10.18653/v1/2022.acl-long.552.

Harvard

Shang, C. et al. (2022) “Improving Time Sensitivity for Question Answering over Temporal Knowledge Graphs”, Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). Association for Computational Linguistics, pp. 8017–8026. Available at: https://doi.org/10.18653/v1/2022.acl-long.552.

Vancouver

1. Shang C, Wang G, Qi P, Huang J (2022) Improving Time Sensitivity for Question Answering over Temporal Knowledge Graphs. In: Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). Association for Computational Linguistics, pp 8017–8026

BibTeX

@inproceedings{shang-etal-2022-improving,
    title = "Improving Time Sensitivity for Question Answering over Temporal Knowledge Graphs",
    author = "Shang, Chao  and
      Wang, Guangtao  and
      Qi, Peng  and
      Huang, Jing",
    editor = "Muresan, Smaranda  and
      Nakov, Preslav  and
      Villavicencio, Aline",
    booktitle = "Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)",
    month = may,
    year = "2022",
    address = "Dublin, Ireland",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2022.acl-long.552/",
    doi = "10.18653/v1/2022.acl-long.552",
    pages = "8017--8026"
}

Metadata:ACL Anthology

Access the Paper

This paper is available from its original source. Click below to access the PDF.

Open PDF

License: https://creativecommons.org/licenses/by/4.0/

Improving Time Sensitivity for Question Answering over Temporal Knowledge Graphs

Abstract

Topics

Keywords

Table of Contents

Knowls

Knowl 1 — Time-Sensitive Question Answering (TSQA) Framework for Temporal Knowledge Graphs

Knowl 2 — Time-Aware Temporal KG Embedding with Auxiliary Time-Order Learning

Knowl 3 — Question Decomposition and Complex Vector Formulation in TSQA

Knowl 4 — Candidate Search Space Pruning via Entity Neighbor Subgraph Extraction

Knowl 5 — Joint Time Estimation and Answer Prediction Scoring Formulation

Knowl 6 — Temporal Contrastive Learning with Dual Time-Order and Answer Guidance