TransE-MTP: A New Representation Learning Method for Knowledge Graph Embedding with Multi-Translation Principles and TransE

Abstract

The purpose of representation learning is to encode the entities and relations in a knowledge graph as low-dimensional and real-valued vectors through machine learning technology. Traditional representation learning methods like TransE, a method which models relationships by interpreting them as translations operating on the low-dimensional embeddings of a graph’s entities, are effective for learning the embeddings of knowledge bases, but struggle to effectively model complex relations like one-to-many, many-to-one, and many-to-many. To overcome the above issues, we introduce a new method for knowledge representation, reasoning, and completion based on multi-translation principles and TransE (TransE-MTP). By defining multiple translation principles (MTPs) for different relation types, such as one-to-one and complex relations like one-to-many, many-to-one, and many-to-many, and combining MTPs with a typical translating-based model for modeling multi-relational data (TransE), the proposed method, TransE-MTP, ensures that multiple optimization objectives can be targeted and optimized during training on complex relations, thereby providing superior prediction performance. We implement a prototype of TransE-MTP to demonstrate its effectiveness at link prediction and triplet classification on two prominent knowledge graph datasets: Freebase and Wordnet. Our experimental results show that the proposed method enhanced the performance of both TransE and knowledge graph embedding by translating on hyperplanes (TransH), which confirms its effectiveness and competitiveness.

1. Introduction

A knowledge graph (KG) serves as a potent tool for structuring knowledge extracted from extensive text and image datasets. Introduced by Google in 2012, the KG is a directed graph comprising nodes and edges, where nodes denote entities or concepts and edges signify various semantic relationships between these entities and concepts. Typical knowledge graph datasets include WordNet [1] and Freebase [2], enabling functionalities such as semantic search, intelligent question answering [3], link prediction [4], and document understanding [5,6]. However, due to the typical scale of knowledge graphs, which normally contain millions of nodes and billions of edges, computational reasoning and operations can be challenging.

For instance, when conducting link prediction on Freebase, we need to deal with 68 million nodes and over a billion edges. Additionally, traditional knowledge graphs rely on symbolic and logical representations, which are not optimal for intensive numerical computations. In recent years, representation learning based on deep learning [7,8] has emerged as a significant paradigm, garnering extensive attention across domains like speech recognition, image processing, and natural language processing. Representation learning focuses on encoding the semantic information of entities and relations into low-dimensional, dense, real-valued vectors with machine learning, facilitating their distributed representation [9]. These distributed representations enable the rapid computation of semantic similarities between entities [10], thereby enhancing computational efficiency, mitigating data sparsity, and integrating heterogeneous information [11].

An overview of recent references found that considering different types of relationships improves modeling effectiveness. In addition, utilizing multi-translation principles for various relationships aids in reasoning and completing more complex knowledge [12]. To address the issue of a single optimization objective within the existing methods, this paper introduces a novel representation learning method, which defines different translation principles for the four types of relationships: one-to-one, one-to-many, many-to-one, and many-to-many [13]. Consequently, entities and their interrelationships can be more effectively represented by establishing specific optimization goals, which enhances the accuracy of predicting missing entities or relationships and facilitates this method’s application to large-scale knowledge graphs [14]. The contributions of our work can be summarized as follows:

• For complex optimization tasks, we first propose multi-translation principles (MTPs). Unlike some existing methods, such as TransE, MTPs can focus on different relationship types and establish corresponding translation principles.
• To produce accurate reasoning and prediction, multiple score functions are constructed based on MTPs for model training. Subsequently, the loss function is optimized with the gradient descent algorithm to minimize the score function.
• Representation learning of a knowledge graph with MTPs aims to optimize multiple objectives by applying translation principles to the four types of relationships: one-to-one, one-to-many, many-to-one, and many-to-many. Furthermore, we introduce TransE-MTP (combining MTPs with TransE). Extensive experiments on two widely used knowledge graph datasets, Freebase and WordNet, demonstrate the efficiency and competitiveness of our proposed model.

The remainder of this paper is organized as follows: In Section 2, work related to knowledge graphs, reasoning and complements for knowledge graphs, etc., are presented. The theory of TransE-MTP is detailed in Section 3. In Section 4, we conduct two subtasks (link prediction and triplet classification) with TransE-MTP, presenting the experimental results and discussing an effectiveness analysis. Finally, Section 5 draws a conclusion and discusses future work.

2. Related Work

2.1. Knowledge Graphs

Knowledge graphs were introduced into search engines by Google Inc. in 2012, sparking a surge in research and applications of large-scale knowledge graphs. Different from a traditional search engine based on a keyword search, knowledge graphs can effectively discover complex related information, comprehend user intent at a semantic level, and enhance query quality [15].

Knowledge representation is essential for acquiring and utilizing knowledge, making it a pivotal aspect of knowledge graph development and application. Knowledge graphs are typically organized as networks, where each node represents an entity (such as a person, place, or concept), and each edge indicates a relationship between entities. Most knowledge can be represented as triplets (h, r, t), where h and t signify the head and tail entities, respectively, and r denotes the relationship between them [16]. This triplet format is standard across all major knowledge graphs, including the resource description framework (RDF) [17] technical standard, which was established by the World Wide Web Consortium (W3C) [18].

2.2. Reasoning and Complement for Knowledge Graphs

Humans possess extensive knowledge bases filled with information, but most of it remains unprocessed. While knowledge graphs are valuable in applications like search engines, they often suffer from incomplete information and errors [19]. Consequently, it is necessary to address these issues in real-world contexts. Reasoning, an essential form of advanced human cognition, involves the process of drawing conclusions based on existing facts and principles [20]. Knowledge reasoning refers to learning how computers or intelligent systems simulate human reasoning with formal knowledge and inference methods. This reasoning capability is a key characteristic of human intelligence, enabling the extraction of hidden knowledge from existing information [21], such as deducing “Obama’s nationality is the United States” from “Obama is the president of the United States,” or determining “grandpa” from “Dad’s dad”. In intelligent systems, the reasoning process is executed by an inference machine, a program specifically designed for reasoning tasks [22].

Knowledge graphs are widely applied for various practical tasks, although their accuracy and completeness cannot always be ensured [23]. To improve this, knowledge graph completion techniques are developed to identify and fill in missing relationships, thereby enhancing the overall quality of the knowledge graph and enabling the development of diverse downstream applications.

The knowledge graph completion task [24] is divided into two subtasks: (1) entity prediction and (2) relationship prediction. Entity prediction involves taking partial triplets (h, r, ?) or (?, r, t) as input (where “?” represents t or h) and predicting an ordered list of candidate entities as output. Relationship prediction focuses on determining the relationship between the head entity and the tail entity. This paper emphasizes the entity prediction task, noting that the methods used for entity prediction can be adapted for relationship prediction by modifying the input.

2.3. Knowledge Representation Learning

Knowledge representation learning embeds the entities and relationships from a knowledge base into vector spaces. It aims to represent the relevant semantic information of research objects as dense, low-dimensional vectors via machine learning techniques [25]. By projecting the entities e and the relations r into a low-dimensional vector space, we obtain vectors le and lr. The similarity between any two objects is then evaluated with metrics such as Euclidean distance. This approach facilitates the efficient expression and computation of the semantic associations between entities and relations [26], addresses the issue of data sparsity, and significantly enhances the performance of knowledge acquisition, fusion, and inference. Furthermore, knowledge representation learning is crucial for the construction, reasoning, and utilization of knowledge graphs.

In recent years, researchers have introduced many models to effectively and rationally represent the entities and relationships in knowledge bases. Notable models include the structural model, single-layer neural network model, energy model, bilinear model, tensor neural network model, matrix analysis model, and translating-based model. The structure embedding (SE) model [27] is a distance-based model that implements link prediction by calculating the closest relationship vector between two entities. The single-layer model (SLM) [28] employs nonlinear operations in a single-layer neural network to improve SE’s semantic representation. The semantic matching energy (SME) model [29,30] captures semantic associations utilizing low-dimensional vectors and multiple projection matrices, though it is computationally intensive. The latent factor model (LFM) [31,32] employs bilinear transformations to capture second-order relationships. The neural tensor network (NTN) [33] enhances traditional neural networks with bilinear tensors, correlating head and tail entity vectors across dimensions. RESCAL [34,35] exemplifies matrix decomposition to derive low-dimensional representation vectors. Since the introduction of word2vec by MiKolov et al. in 2013 [36], representation learning has become vital in natural language processing (NLP) for identifying imbalances in word vector spaces. Inspired by this, TransE [37] maps entities and relations into a low-dimensional space, modeling relations as translations between entity embeddings. To address TransE’s limitations, knowledge graph embedding by translating on hyperplanes (TransH) [38] allows entities to have varied representations under different relationships. Learning entity and relation embeddings for knowledge graph completion (TransR) [39] projects each triplet into its respective relational space for translation. Knowledge graph embedding via a dynamic mapping matrix (TransD) [40] utilizes two projection matrices to map entities into relational space, reducing parameter overflow. Knowledge graph completion with an adaptive sparse transfer matrix (TranSparse) [41] replaces the dense matrices with sparse matrices to address heterogeneity and imbalance. An adaptive approach for knowledge graph embedding (TransA) [42] employs Euclidean distances for loss computation and vector weighting, improving prediction accuracy and efficiency. Knowledge representation learning is crucial for deriving vector representations of semantic information and enabling the effective computation of entities, relationships, and their interactions [43].

3. The Proposed Method

In this paper, we propose a new model of representation learning for the knowledge graph embedding multi-translation principle and TransE, named TransE-MTP. As previously discussed, TransE is a representative model for knowledge representation learning. It is an energy-based model (also known as a translating-based model). Figure 1 illustrates the TransE framework, showing the vector positions of the head and tail entities along with their relationships. The horizontal and vertical axes denote the distances from the origin of the head entity (h), tail entity (t) and relation (r) [44]. In Figure 1, h, r, and t represent the head entity vector, relation vector, and tail entity vector, respectively.

The fundamental idea of TransE is to embed all entities and relations from the knowledge graph into a unified low-dimensional vector space [45]. The loss function, optimized with the gradient descent algorithm, aims to minimize the score function to achieve h + r = t. The score function, defined as Equation (1), is designed to be lower for positive triplets and higher for negative triplets:

$$f_r\left(h,t\right)={‖h+r-t‖}_{L_{1/2}}^2$$

(1)

TransE has achieved an efficient prediction performance by embedding large-scale knowledge graphs of entities and relationships into a low-dimensional vector space. However, it struggles with complex relationships like reflexivity, one-to-many, many-to-one, and many-to-many due to it ignoring the distributed representation of entities within different relationships [46]. For instance, in Figure 2, a one-to-many relationship such as (Song dynasty, poet, Shi Su), (Song dynasty, poet, You Lu), …, (Song dynasty, poet, …), (Song dynasty, poet, Qiji Xin) may lead TransE to incorrectly predict a result like Shi Su = You Lu = … = Qiji Xin because there are many poets in the Song dynasty.

The existing TransE-based models typically apply a single translation principle to all types of relations. However, each type of relation should have its own optimization objective. To address this, our method TransE-MTP defines distinct translation principles for different types of relations, as detailed below:

• For one-to-one and many-to-one relations, the translation principle is defined as h + r = t;
• For one-to-many relations, the translation principle is defined as t − r = h;
• For many-to-many relations, the translation principle is defined as t − h = r.

In this study, we first combine the MTPs with TransE and conduct experiments on their link prediction and triplet classification using Freebase and Wordnet. Then, we compare our experimental results with those of TransE, TransH, and several other previous models of representation learning, analyzing and evaluating their performance. The results indicate that TransE-MTP exhibits significant advantages in both link prediction and triplet classification tasks, outperforming previous models considerably. Therefore, defining different translation principles by distinguishing relational mapping types can not only enhance the accuracy of representing entities, relations, and their interconnections, but also prove beneficial for large-scale knowledge graphs.

Numerous knowledge representation learning models have been developed to learn the representations of entities and relations in knowledge bases. Existing translation-based representation learning methods aim to embed the entities and relations of knowledge graphs into a continuous vector space in the form of vectors [47], and typically evaluate them using a defined loss function. Various embedding models employ different loss function definitions for the corresponding triplets in the knowledge graph. The method proposed in this paper addresses the single optimization objective issue present in the existing methods and effectively applies it to large-scale knowledge graphs.

To better understand these models, we must first define some common symbols. Here, h and t denote the head entity and tail entity, respectively, while r represents the relation between them. The optimal value of the loss function is obtained by minimizing the gap between a positive and negative triplet through the stochastic gradient descent (SGD) algorithm, typically represented as in Equation (2).

$$L={\displaystyle {\sum }_{\left(h,r,t\right)}{\displaystyle {\sum }_{\left(h^{\prime },r^{\prime },t^{\prime }\right)}{\left[f\left(h,r,t\right)+\gamma -f\left(h^{\prime },r^{\prime },t^{\prime }\right)\right]}_{+}}}$$

(2)

Here, (h, r, t) belongs to the set S(h, r, t) of all positive triplets in the knowledge base, and (h′, r′, t′) belongs to the set S′(h, r, t) of negative triplets. The threshold value γ distinguishes positive from negative triplets. Negative triplets are generated by replacing one of the head, relation, and tail entities of each positive triplet in S with another entity or relation, resulting in the negative triplet set S′ denoted as {(h′, r, t), (h, r, t), (h, r, t′)}.The basic knowledge of a knowledge graph is introduced as follows: G = (E, R, S) represents a knowledge base, where E denotes the set of all entities in the knowledge base, including E different kinds of entities, and R denotes the set of all relations, including R different kinds of relations. γ is a margin which can distinguish the distance between positive triplets and negative triplets. Triplets are generally represented as (h, r, t), where h and t as the head entity and tail entity, respectively, and r denotes the relation between the head and tail entities. For instance, the triplet (Ming Yao, birthplace, Shanghai) indicates that there is a “birthplace” relation between “Ming Yao” and “Shanghai”.

In this paper, we present TransE-MTP by combining MTPs with TransE. TransE embeds entities and relations into a low-dimensional vector space, learning the embedding vectors for entities and relations from a training set S of triplets (h, r, t), composed of two entities h and t (belonging to the entity set E) and a relation r (belonging to the relation set R). TransE primarily learns to represent each relation as a vector equation h + r = t, optimizing it through training. To advance knowledge representation, TransE-MTP defines different translation principles for different relation types and calculates optimal values accordingly [37]. Subsequently, we perform link prediction and triplet classification on the relevant datasets.

4. Experiment and Analysis

4.1. Dataset and Experimental Setup

In this paper, we conduct link prediction and triplet classification on two prominent knowledge graph datasets: WordNet and Freebase. WordNet is a comprehensive lexical knowledge graph that offers semantic knowledge about words. The entities in Wordnet represent the synonyms of various concepts, while the relations capture the lexical relationships between these semantic concepts [48]. Specifically, we work with two subsets of WordNet, WN11 and WN18, which include 11 and 18 types of relations, respectively [49]. Freebase, another large-scale knowledge graph, holds a vast number of facts, comprising approximately 1.2 billion triplets and 80 million entities. We utilize two subsets of Freebase, FB15K and FB13, with their statistics detailed in

Table 1. Statistics of the datasets.

Dataset	#Rel	#Ent	#Train	#Valid	#Test
WN11	11	38,696	112,581	2609	10,544
WN18	18	40,493	141,442	5000	5000
FB13	13	75,043	316,232	5908	23,733
FB15K	1345	14,951	483,142	50,000	59,071

. In Table 1, the column #Rel represents the number of relational types in the corresponding dataset, and #Ent denotes the number of entities. #Train, #Valid, and #Test represent the number of triplets corresponding to the training set, verification set, and test set, respectively.

Furthermore,

Table 2. Parameters and complexity analysis.

Model	#Parameters	#Operations (Time Complexity)
SLM	$O\left(N_{e}m+N_r\left(2k+2nk\right)\right)\left(m=n\right)$	$O\left(\left(2mk+k\right)N_t\right)$
NTN	$O\left(N_{e}m+N_r\left(n^{2}s+2ns+2s\right)\right)\left(m=n\right)$	O (((m2 + m) + 2mk+ k)Nt)
TransE	$O\left(N_{e}m+2N_{r}n\right)\left(m=n\right)$	$O\left(N_t\right)$
TransH	$O\left(N_{e}m+2N_{r}n\right)\left(m=n\right)$	$O\left(2m{N}_t\right)$
TransE-MTP	$O\left(N_{e}m+2N_{r}n\right)\left(m=n\right)$	$O\left(N_t\right)$

outlines the complexity of all the models discussed in this paper, and we compare these complexities with that of TransE-MTP. This comparison considers the complexity of the embedded model, including the number of parameters and the number of multiple operations per iteration. Here, N_e and N_r represent the number of entities and relations; N_t stands for the number of triplets in the knowledge graph; m indicates the dimension of the entity embedding space; n denotes the dimension of the relation embedding space; d represents the number of clusters for a relationship; k is the number of implicit nodes in a neural network; s denotes the number of slices of the tensor; and $\tilde{\theta }(0\le \tilde{\theta }\le 1)$ signifies the average sparsity of all translation matrices. As illustrated in Table 2, the complexity of TransE-MTP is lower than that of TransH and is the same as TransE, which indicates that TransE-MTP does not increase the complexity and even has a lower complexity than other models.

4.2. Link Prediction

Link prediction focuses on identifying the missing entity h (or t) in a fact triplet (h, r, t). Instead of finding a single optimal answer, it emphasizes ranking the sets of candidate entities derived from the knowledge graph. In this paper, we present two datasets: WN18, a subset of Wordnet, and FB15K, a relatively dense subgraph in Freebase, with all entities available in the Wikipedia database.

Firstly, we replace the head entities or tail entities with all entities in the knowledge graph for each triplet (h, r, t) that needs to be tested, while calculating the similarity score of these entities through the score function f. We then sort these entities in descending order according to their similarity score. Meanwhile, we employ two evaluation metrics: (1) the average ranking of correct entities (known as the Mean Rank) and (2) the proportion of correct entities ranked in the top 10 ratio (known as the Hit@10). Obviously, an outstanding model should perform well at both Mean Rank and Hit@10.

It is important to note that negative triplets might exist in the knowledge graph and could be mistakenly treated as positive triplets. This could lead to an underestimation of the model’s performance if these negative triplets are ranked higher than the positive ones. Therefore, we should filter out these triplets from the training set, validation set, and test set before ranking. We refer to this filtered evaluation setting as “Filt” and the original evaluation metric setting as “Raw”. Regarding the strategy of constructing negative labels, “unif” represents the traditional method of replacing the head or tail with one of an equal probability, and “bern” denotes reducing false negative labels by replacing the head or tail with ones of different probabilities.

Table 3. TransE-MTP parameter statistics for link prediction.

Dataset	Model	Λ	γ	n, m	B	D.S
WN18	TransE-MTP	0.01	3.5	50	1440	L1
FB15K	TransE-MTP	0.001	1	200	4800	L1

lists the training parameters utilized for TransE-MTP in our experiment. These values represent the optimal parameters, allowing the experiment to yield the best outcomes. Specifically, λ denotes the learning rate, γ represents the margin, n and m indicate the embedding dimensions for the entity and relation, B is the size of the mini-batch, and D.S stands for the dissimilarity measured in the score function. The iteration number for the SGD is 1000.

Table 4. Results of link prediction.

	Dataset	WN18				FB15K
Model
Metric		Mean Rank		Hits@10		Mean Rank		Hits@10
		Raw	Filt	Raw	Filt	Raw	Filt	Raw	Filt
RESCAL		1180	1163	37.2	52.8	828	683	28.4	44.1
SE		1011	985	68.5	80.5	273	162	28.8	39.8
SME		542/526	533/509	65.1/54.7	74.1/61.3	274/284	154/158	30.7/31.3	40.8/41.3
LFM		469	456	71.4	81.6	283	164	26.0	33.1
TransE		263	251	75.4	89.2	243	125	34.9	47.1
TransH (unif/bern)		318/401	303/388	75.4/73.0	86.7/82.3	211/212	84/87	42.5/45.7	58.5/64.4
TransE-MTP (unif/bern)		261/238	249/226	73.2/76.6	85.2/87.3	237/241	83/160	50.8/52.5	78.6/75.2

lists the results of the link prediction; compared to the existing RESCAL, SE, SME, LFM, TransE and TransH, TransE-MTP surpasses most traditional representation learning models on WN18 and FB15K. TransE-MTP outperforms TransE by improving the Mean Rank (Raw and Filt) by 25 and the Hits@10 (Raw) by 3.4% on WN18, and by improving the Mean Rank (Filt) by 42, the Hits@10 (Raw) by 15.9%, and the Hits@10 (Filt) by 21.5% on FB15K. Compared with TransH, TransE-MTP improves the Mean Rank (Raw) by 163, the Mean Rank (Filt) by 162, the Hits@10 (Raw) by 3.6%, and the Hits@10 (Filt) by 5% on WN18. Meanwhile, it improves the Mean Rank (Filt) by 1, the Hits@10 (Raw) by 8.3%, and the Hits@10 (Filt) by 20.1% on WN11. Overall, the experimental results show significant improvement, with TransE-MTP’s Mean Rank on WN18 outperforming most traditional representation learning models. The results of the link prediction also show that TransE-MTP is more effective in its reasoning and completion of knowledge graphs and can be applied to functionalities such as semantic searches, intelligent question answering, recommendation systems, document understanding, etc.

4.3. Triplet Classification

The purpose of triplet classification is to determine whether a given triplet is positive, which is a binary classification task. In alignment with existing models in the literature, we utilized three datasets, WN11, FB13, and FB15K, for our experiment on triplet classification. Additionally, negative triplets are generated on FB15K by corrupting positive triplets with the same settings. For triplet classification, a threshold ${\delta }_r$ is established to optimize classification accuracy on the verification set. The classification criterion is as follows: if the triplet’s dissimilarity is less than ${\delta }_r$, it is classified as a positive triplet; otherwise, it is classified as a negative triplet. ${\delta }_r$ is determined by maximizing the classification accuracy on the verification set.

Table 5. TransE-MTP parameter statistics for triplet classification.

Dataset	Model	α	λ	n, m	B	D.S
WN11	TransE-MTP	0.01	4	20	120	L1
FB13	TransE-MTP	0.01	2	100	480	L1
FB15K	TransE-MTP	0.001	1	200	4800	L1

presents the parameters for training TransE-MTP with the optimal parameter values.

Table 6. Results of triplet classification.

	Dataset	WN11	FB13	FB15K
Model
SE		53.0	75.2	-
SME (bilinear)		70.0	63.7	-
SLM		69.9	85.3	-
LFM		73.8	84.3	-
NTN		70.4	87.1	68.2
TransE (unif/bern)		75.9/75.9	70.9/81.5	77.3/79.8
TransH (unif/bern)		77.7/78.8	76.5/83.3	74.2/79.9
TransE-MTP (unif/bern)		81.7/81.3	79.8/84.86	84.2/83.6

shows the results of the triplet classification compared to the existing RESCAL, SE, SME, LFM, TransE, and TransH models, indicating that TransE-MTP significantly outperforms TransE. Specifically, TransE-MTP achieves accuracy rates of 81.7%, 84.86%, and 84.2% on WN11, FB13, and FB15K, respectively. Notably, TransE-MTP surpasses TransE, achieving a 5.4% higher accuracy on WN11 (bern), 8.9% on FB13 (unif), and 6.9% on FB15K (unif). Compared to TransH, TransE-MTP improves the accuracy by 4% on WN11 (unif), 3.3% on FB13 (unif), and 10% on FB15K (unif), demonstrating a substantial improvement. The results of the triplet classification suggest that TransE-MTP is more accurate in triplet classification and represents the knowledge graph more effectively. The experimental results mentioned above are directly derived from references, as all methods discussed in this paper employ the same testing dataset. Clearly, a proficient model should achieve a high triplet classification accuracy as TransE, TransH, and TransE-MTP have.

5. Conclusions and Future Work

This paper introduces TransE-MTP by combining MTPs with TransE, which mainly involves the following work: (a) defining distinct translation principles for different relation types, such as one-to-one and complex relations like one-to-many, many-to-one, and many-to-many, which can enhance the capacity of knowledge graphs’ reasoning and completion; (b) combining MTPs with TransE, which not only maps the entities and relations into low-dimensional real-valued vector spaces, but also addresses the issue of a single optimization objective; and (c) conducting experiments on link prediction and triplet classification on the FB15K, WN11, FB13, and WN18 datasets. Extensive experimental results demonstrate that TransE-MTP improves the performance of both link prediction and triplet classification tasks. It distinguishes between different relation types, clarifies objects during training, and improves prediction accuracy without adding complexity. Moreover, it significantly enhances the capacity for automatic knowledge acquisition, fusion, and reasoning, facilitating more extensive and in-depth applications of knowledge graphs based on distributed representation learning.

In future research, while continuing to define different translation principles for distinct relation types, there will be growing focus on multi-information fusion and complex relational path reasoning to address the key challenges in knowledge graph representation learning, such as the precision of knowledge graph embedding, computational efficiency, and data sparsity. We will also explore the integration of textual, pictorial, and structured information, focusing on the reasoning of complex relational paths and the completion of knowledge graphs. Meanwhile, the proposed method is promising for extending to more applications such as semantic searches, intelligent question answering, recommendation systems, and document understanding. Additionally, we can utilize knowledge representation learning methods to analyze ancient Chinese poetry’s implications for climate studies, which is highly significant for the current research on climate change and global warming.