Knowledge Graph-Based Question Answering via Enhanced Contrastive Learning

1. Introduction

Dialog-driven query resolution (DDQR) using knowledge graphs (KGs) is a crucial, evolving field, increasingly relevant due to the rise of sophisticated personal assistant technologies like Siri and Google Assistant. Current literature in this domain generally diverges into two methodologies [12,43,45]: semantic parsing, which transforms user queries into executable logical forms on KGs, and information retrieval, which focuses on optimizing graph path selection based on query specifics [23,48].

Question Answering (QA) [57] systems are designed to retrieve information or provide answers to questions posed in natural language. The development of QA systems dates back to the early days of artificial intelligence, with initial systems focusing on structured data retrieval from curated databases. Over the years, the evolution of QA has been influenced by advancements in computational linguistics and machine learning, enabling these systems to parse complex queries and fetch relevant answers from large, unstructured datasets. Modern QA systems, such as those integrated into search engines and virtual assistants, employ sophisticated natural language processing techniques to understand the intent and context of a question, making them integral tools in information retrieval and decision support [60].

In the contemporary landscape, QA systems have diversified into various sub-fields, including factoid QA, list QA, definition QA, and conversational QA [57,58]. Each type caters to different user needs; for example, factoid QA focuses on providing concise, factual answers to specific questions, while conversational QA aims to engage the user in a dialogue, offering responses that require contextual understanding over multiple turns. This segmentation has prompted significant research in domain-specific QA systems, where knowledge bases are tailored to specific fields such as medicine, finance, or law. The integration of machine learning models, especially deep learning, has further propelled the capabilities of QA systems, enabling them to learn from vast amounts of data and improve their accuracy over time through techniques such as reinforcement learning and transfer learning.

The challenge with semantic parsing is its intensive demand for annotated training data [23], each query needing a precise logical counterpart. Conversely, retrieval-focused approaches merely require correct end entities or paths per query, simplifying data preparation but complicating the retrieval process.

Existing methods in DDQR often categorize the task as path optimization. Solutions range from heuristic-based models [10,67] to sophisticated reinforcement learning schemes [21]. However, these models either demand extensive manual rule crafting or heavily rely on user engagement for query refinement, which can burden users and lead to dissatisfaction if frequent adjustments are needed within short intervals [14].

This work introduces a novel framework, CONVEX (CONVersational EXploration), which leverages contrastive learning to refine KG path ranking by effectively incorporating dialog histories and domain insights into the learning process. The choice for contrastive learning stems from its capacity to differentiate between relevant and irrelevant paths by contrasting them against each other in a shared embedding space [34].

By integrating entire dialog histories and domain information, CONVEX aims to enhance the representational depth, improving the system’s ability to discern and prioritize the most relevant KG paths. This approach addresses significant limitations in existing models by reducing dependency on manual rule creation and mitigating the cognitive load on users [69,70].

Key contributions of this paper are: 1) Introduction of CONVEX, a pioneering contrastive learning model for DDQR, designed to cohesively blend conversational context with KG path learning to elevate path ranking efficacy. 2) An extensive analysis of how additional contextual layers, such as dialog domain and response fluency, impact the performance of DDQR systems. Our findings indicate substantial improvements over traditional methods on several benchmark datasets. 3) Detailed experimental insights that validate the effectiveness of CONVEX, showcasing significant enhancements in ranking metrics across multiple scenarios.

The remainder of the paper is organized as follows: Section 2 reviews related work. Section 3 outlines the task definitions and notations. Section 4 describes the CONVEX framework. Section 5 discusses experimental setups, results, and both ablation studies and error analyses. We conclude our findings in Section 6.

2. Related Work

Question Answering over Knowledge Graphs (KGQA) represents a robust area of inquiry within the broader field of artificial intelligence. Comprehensive surveys on this topic have categorized various approaches, with specific details found in foundational works by Zaib et al. [43] and Diefenbach et al. [12], Feietal. [74]. Our research intersects with these established methodologies but introduces novel elements that are particularly relevant to conversational contexts.

Historically, KGQA efforts have emphasized the development of semantic graph generation and subsequent re-ranking processes. Initial strategies involved comparing user queries against predefined query templates, as demonstrated by Bast and Haussmann [1], Fei et al. [76], who enriched these templates with potential relations to create a variety of query graph candidates. Further refinement was introduced by Yih et al.[41], who implemented a heuristic search algorithm staged to construct grounded query graphs, which were then evaluated using a neural ranking model to select the most accurate semantic graph. This paradigm was extended by Yu et al.[42] through the integration of a hierarchical representation of KG relations into a neural query graph ranking model, which was compared against a local sub-sequence alignment model employing cross-attention mechanisms [32]. Maheshwari et al.[28], Fei et al. [82] further explored this area by conducting a comprehensive empirical investigation of various neural query graph ranking models, culminating in the development of a self-attention-based slot matching model that leverages the intrinsic structure of query graphs for enhanced performance.

In the domain of Conversational Question Answering (ConvQA), seminal works have largely adopted semantic parsing techniques to tackle multi-turn dialogues [17,33,38,85]. The pioneering model by Saha et al. [36] combined elements of the Hierarchical Recurrent Encoder-Decoder (HRED) model [37] and the key-value memory network model [30], integrating these components to effectively compute contextual representations and facilitate dynamic answer generation. This approach was pivotal in establishing a framework that could accommodate the complexities inherent in conversational interactions, setting the stage for subsequent advancements. Follow-up studies have built on this foundation, exploring multi-task learning frameworks that further refine the handling of conversational nuances and query adjustments [17,33].

Christmann et al. [10] and Kaiser etal. [21] introduced innovative methods for maintaining conversational context and adapting to user interactions within a KG. These methodologies, while groundbreaking, exhibited limitations related to their reliance on rule-based systems or user-driven query reformulation, which can impose significant cognitive demands on users and are susceptible to biases and errors. Our work, named CONVEX, seeks to transcend these limitations by adopting a weak supervision approach that harnesses the full spectrum of conversational context and available KG paths to retrieve answers, eschewing the need for direct user input in the reformulation process.

Contrastive learning, initially introduced for facial recognition [9], has seen widespread application across various domains, including image caption identification [35], augmented image code computation [5], feature clustering [4], and dense information retrieval [15]. CONVEX adapts this framework to the domain of KGQA by developing a contrastive learning model that minimizes the embedding distance between similar items and maximizes it for dissimilar ones within the context of conversational interactions, fluent responses, and domain-specific information [93]. The subsequent sections will delve into the specifics of CONVEX’s methodology, illustrating how it addresses the unique challenges of conversational question answering over knowledge graphs.

3. Preliminary

In the realm of Knowledge Graphs (KGs), a KG can be described as a triplet $\mathcal{K}=(\mathcal{E},\mathcal{R},{\mathcal{T}}^{+})$, where $\mathcal{E}$ represents the set of entities, $\mathcal{R}$ denotes the relations, and ${\mathcal{T}}^{+}$ consists of triples indicating relational connections between entities. Specifically, a triple $\tau =(e_h,r_{h,t},e_t)\in {\mathcal{T}}^{+}$ asserts a relationship $r_{h,t}$ linking a head entity $e_h$ to a tail entity $e_t$. In the context of dialog-driven query answering, a conversation $\mathcal{C}$ spanning $\mathrm{T}$ turns is framed by sequences of questions $\mathcal{Q}=\left\{q^t\right\}$ and corresponding answers $\mathcal{A}=\left\{a^t\right\}$, structured as $\mathcal{C}=\langle (q^0,a^0),(q^1,a^1),...,(q^{\mathrm{T}},a^{\mathrm{T}})\rangle$. Each question $q^t$ and its fluent response $v^t$ are token sequences, enhancing the conversational context and enabling richer interaction dynamics within the KG framework.

The conversation context is enriched by defining:

• Context Entities ${\mathcal{E}}_c$: A subset of $\mathcal{E}$, including entities mentioned throughout $\mathcal{C}$.
• Context Paths ${\mathcal{P}}_c$: Paths derived from ${\mathcal{E}}_c$, spanning up to three hops, thus encapsulating the breadth of relational connectivity relevant to the conversational queries.

Problem Formulation: The ConvQA task under our newly proposed CONVEX model aims to utilize the KG $\mathcal{K}$, integrating the conversational history ${\mathcal{C}}^t$ and context entities ${\mathcal{E}}_c^t$ to identify viable KG paths ${\mathcal{P}}_c^t$. This process is approached as an information retrieval (IR) challenge, scoring and ranking potential paths to ascertain those most likely to yield correct responses as specified by gold standard answers $a^t$. The CONVEX model employs contrastive representation learning to optimize this selection, ensuring that similar contexts yield proximal embeddings, while distinct scenarios are clearly separated in the embedding space.

Table 1. Notational Glossary for the CONVEX Framework.

Notation	Definition
$\mathcal{K},\mathcal{E},\mathcal{R},{\mathcal{T}}^{+}$	Knowledge Graph, Entities, Relations, Triples
$\mathcal{C},t$	Conversation, Turn Index
$q^t,a^t$	Query and its Answer at turn t
$v^t$	Verbatim response at turn t
${\tau }^t$	Tail entity at turn t
${\mathcal{C}}^t$	Historical Context up to turn t
${\mathcal{E}}_c,{\mathcal{P}}_c$	Contextual Entities, Derivable Paths
${\mathcal{D}}^{t+},{\mathcal{D}}^{t-}$	Positive/Negative Path Sets for $q^t$
$s^t$	Input sequence incorporating ${\mathcal{C}}^t$ and $q^t$
d	Dimensionality of the embedding space
$h^{(·)}$	Contextual embeddings function
${\theta }^{(·)}$	Parameters subject to optimization
${\mathbf{W}}^{(·)}$	Weight matrices in neural layers
${\omega }^{(·)}$	Probability distributions over output space
${\varphi }^c,{\varphi }^p$	Joint embeddings for conversation and path contexts

Table 1. Notational Glossary for the CONVEX Framework.

Notation	Definition
$\mathcal{K},\mathcal{E},\mathcal{R},{\mathcal{T}}^{+}$	Knowledge Graph, Entities, Relations, Triples
$\mathcal{C},t$	Conversation, Turn Index
$q^t,a^t$	Query and its Answer at turn t
$v^t$	Verbatim response at turn t
${\tau }^t$	Tail entity at turn t
${\mathcal{C}}^t$	Historical Context up to turn t
${\mathcal{E}}_c,{\mathcal{P}}_c$	Contextual Entities, Derivable Paths
${\mathcal{D}}^{t+},{\mathcal{D}}^{t-}$	Positive/Negative Path Sets for $q^t$
$s^t$	Input sequence incorporating ${\mathcal{C}}^t$ and $q^t$
d	Dimensionality of the embedding space
$h^{(·)}$	Contextual embeddings function
${\theta }^{(·)}$	Parameters subject to optimization
${\mathbf{W}}^{(·)}$	Weight matrices in neural layers
${\omega }^{(·)}$	Probability distributions over output space
${\varphi }^c,{\varphi }^p$	Joint embeddings for conversation and path contexts

4. Methodology

Our methodology, designated as CONVEX, employs a contrastive learning framework to effectively rank knowledge graph (KG) paths based on the dialogue context of each conversational turn. The process, depicted is structured into three primary phases: 1) preprocessing to identify potential KG paths, 2) encoding the conversational context, and 3) embedding the dialogue and KG paths jointly for contrastive path ranking.

Preprocessing & Path Extraction: Initially, we process the input comprising questions $q^t$ and answers $a^t$ to identify relevant context entities and potential KG paths ${\mathcal{P}}_c^t$, similar to the approach used in [21]. These paths are preliminarily scored based on their alignment with the domain-specific characteristics of the query. The embeddings for these paths are prepared using sentence transformers that leverage BERT’s [11] deep contextual representations.

Contextual Encoding: We further enrich the model by encoding the conversational history and integrating domain knowledge. A BART-based bidirectional encoder [25] captures the nuances of both the historical and current conversational context, generating embeddings that reflect the sequential interplay of dialogue exchanges. This step is crucial for aligning the conversational flow with the structured knowledge embedded in KG paths.

Joint Embedding and Ranking: The core of CONVEX lies in its ability to jointly embed the dialogue context and the KG paths within a shared embedding space. This shared space facilitates a contrastive ranking mechanism, where the goal is to minimize the distance between correct path-context pairs and maximize the distance for incorrect pairs. This mechanism is critical for enhancing the accuracy of the model in real-time conversational settings.

In more detail, the preprocessing phase involves identifying and encoding potential KG paths using domain-specific transformers that prepare the path representations for subsequent ranking tasks. These representations are tuned to capture the essential features of the paths that are most likely to answer the given query in a conversation.

The contextual encoding phase leverages a sophisticated encoder to integrate the conversation history, which includes both the questions and the corresponding fluent responses. This integrated approach helps in understanding the progression of the conversation, thus providing a richer context for the subsequent ranking of KG paths.

The joint embedding and ranking phase is designed to operate in an end-to-end fashion post-preprocessing. It intricately combines the conversation’s encoded context with the candidate KG paths, employing a contrastive learning strategy to distinguish between the most and least relevant paths effectively. This phase is pivotal in ensuring that the ranked paths are not only contextually relevant but also dynamically aligned with the ongoing conversation.

As illustrated in Table 2, each step of the CONVEX methodology utilizes cutting-edge AI techniques to enhance the interaction between conversational context and KG path ranking, ensuring a robust and adaptive QA system. The implications of such an approach are profound, allowing for more nuanced and accurate responses within conversational AI applications.

4.1. Joint Contrastive Learning

CONVEX integrates three core trainable modules, each associated with a specific loss function. These modules include the encoder, domain identification pointer, and contrastive ranking module, each critical for the joint learning process. The losses from these modules are combined in a weighted average to optimize the model holistically:

$$L={\lambda }_1{L}^{dm}+{\lambda }_2{L}^{rk}+{\lambda }_3{L}^{dec},$$

(1)

where ${\lambda }_1,{\lambda }_2,{\lambda }_3$ are the weighting factors for each loss component. The loss components are defined as:

$$\begin{array}{cc}\hfill & L^{dm}=-\sum _{j=1}^{m}logp\left(y_j^{\left(dm\right)}\right|x),\hfill \\ \hfill & L^{rk}=\left\{\begin{array}{cc}1-cos({\varphi }^c,{\varphi }^p),\hfill & \mathrm{if}{y}^{\left(rk\right)}=1\hfill \\ max(0,cos({\varphi }^c,{\varphi }^p)-\alpha ),\hfill & \mathrm{if}{y}^{\left(rk\right)}=-1\hfill \end{array}\right.,\hfill \\ \hfill & L^{dec}=-\sum _{l=1}^{n}logp\left(y_l^{\left(dec\right)}\right|x),\hfill \end{array}$$

(2)

These definitions reflect the integration of various learning signals to improve the model’s effectiveness across different tasks. The complete learning algorithm for CONVEX is outlined in Algorithm 1.

Algorithm 1:CONVEX Algorithmic Implementation

Input: Training set $S_{train}=\left\{(q^t,C^t,v^t,{\tau }^t,{\mathcal{D}}_c^{t+},{\mathcal{D}}_c^{t-})\right\}$

4.2. Inference Process

Post-training, CONVEX’s inference involves several steps to utilize the learned models effectively. Initially, the context entities ${\mathcal{E}}^c$ are identified within the input question and conversational history. Subsequently, relevant context paths ${\mathcal{P}}^c$ are extracted, and the conversation is encoded to identify the domain and score the candidate paths. The highest-ranked path determines the response, which is then used to generate a fluent answer via the decoder module.

5. Experiments

5.1. Settings

Model Implementations.Table 3 summarizes the hyperparameters used in the experiments. The CONVEX framework is implemented with a uniform space dimension of $d=768$ across all components. The encoder and decoder are built upon the BART (base) architecture [25]. Training parameters include a batch size of 32, a learning rate of $1e-4$, over 120 epochs, with model checkpoints saved periodically. Optimization is conducted using the AdamW algorithm, with a specific adjustment for weight decay [27]. We apply a residual dropout of $0.1$ across various modules, including the domain pointer and the ranking module, to mitigate overfitting. Initial embeddings for domains and KG paths are generated using a pre-trained BERT model. The maximum token limit for the input sequence combining conversational history and the current question ($C^t+q^t$) is set to 150. We apply weighting factors ${\lambda }_1=0.25$ and ${\lambda }_2=0.25$ for the domain pointer and decoder losses, respectively, and ${\lambda }_3=1.0$ with a margin $\alpha =0.1$ for the path ranking cosine embedding loss.

Datasets and Models for Comparison. For evaluating the performance of CONVEX, we utilize the ConvQuestions [10] and ConvRef [21] datasets, incorporating fluent responses [19]. Baselines include the original CONVEX model [10] and CONQUER [21], an RL-based conversational QA method. We also compare against recent semantic parsing-based model OAT [29] and Focal Entity [24].

Evaluation Metrics. The effectiveness of CONVEX is assessed using Precision at the top rank (P@1), Mean Reciprocal Rank (MRR), and Hit at 5 (H@5), along with precision, recall, and F1-score for domain identification. BLEU-4 and METEOR scores are utilized for evaluating fluent response generation.

5.2. Results

Research Questions: The primary research question, RQ, asks: What is the efficacy of the contrastive learning approach in CONVEX for ranking KG paths? This question is further subdivided into: RQ1.1: What impact does conversational context have on CONVEX’s efficiency? RQ1.2: How do individual tasks within CONVEX perform, such as domain identification and fluent response generation?

Overall Performance on ConvQA datasets.Table 4 shows that CONVEX surpasses all previous baselines on the ConvQuestions dataset in metrics such as P@1, H@5, and MRR. Similarly, for ConvRef, CONVEX outperforms existing models, particularly in metrics that evaluate deeper conversational understanding and context utilization.

5.3. Ablation Study

We perform an ablation study to delineate the contributions of different components in CONVEX. Removing the full conversational context, domain information, or fluent responses each detrimentally impacts the performance, underscoring their collective importance in our model’s architecture.

Table 5. Ablation study results showing the impact of different components and training strategies on the performance of CONVEX.

Configuration	ConvQuestions			ConvRef
Full Model	0.292	0.529	0.398	0.335	0.599	0.441
w/o Full Conv. History	0.214	0.375	0.299	0.247	0.449	0.324
w/o Domain Info	0.247	0.436	0.296	0.266	0.472	0.356
w/o Fluent Responses	0.265	0.441	0.324	0.279	0.503	0.397
Train Separately	0.255	0.413	0.328	0.304	0.529	0.408

Table 5. Ablation study results showing the impact of different components and training strategies on the performance of CONVEX.

Configuration	ConvQuestions			ConvRef
Full Model	0.292	0.529	0.398	0.335	0.599	0.441
w/o Full Conv. History	0.214	0.375	0.299	0.247	0.449	0.324
w/o Domain Info	0.247	0.436	0.296	0.266	0.472	0.356
w/o Fluent Responses	0.265	0.441	0.324	0.279	0.503	0.397
Train Separately	0.255	0.413	0.328	0.304	0.529	0.408

5.4. Error Analysis

Our detailed error analysis, focusing on 250 randomly sampled incorrect predictions, highlights two main types of errors: Incorrect ranking of paths with semantically similar KG relations and the absence of complete gold KG paths in the training data, underscoring areas for potential improvement in future iterations of CONVEX.

Table 6. Summary of performance improvements achieved by CONVEX over the baseline.

Metric	Baseline	CONVEX	Improvement
Hit@5	0.343	0.529	+54.2%
Precision@1	0.240	0.292	+21.7%
MRR	0.279	0.398	+42.7%

Table 6. Summary of performance improvements achieved by CONVEX over the baseline.

Metric	Baseline	CONVEX	Improvement
Hit@5	0.343	0.529	+54.2%
Precision@1	0.240	0.292	+21.7%
MRR	0.279	0.398	+42.7%

5.4.1. Incorrect Ranking of Semantically Similar KG Relations

CONVEX occasionally misranks paths when they contain KG relations that are semantically similar but contextually distinct. Consider the query, "What kind of book is it?" set within a rich conversational history:

• $q^1$: "Who wrote The Secret Garden?"
• $v^1$: "Frances Eliza Hodgson Burnett."
• $q^2$: "Where is the story set?"
• $v^2$: "In Yorkshire."
• $q^3$: "When was it published?"
• $v^3$: "In 1910."

The system needs to identify the KG path containing the relation "main subject (P921)," pointing correctly to "adventure (Q1436734)." However, CONVEX often prioritizes paths featuring the "genre (P136)" relation due to its prevalent usage across training instances. This example underscores a key challenge: distinguishing between closely related semantic relations like "main subject" and "genre," which often leads to suboptimal ranking outcomes.

5.4.2. Absence of Gold KG Paths

A significant challenge in our dataset is the absence of gold-standard KG paths, affecting over 25% of the training examples and 19% of the test cases. This absence has a direct impact on CONVEX’s learning efficacy, as instances lacking a gold path are invariably treated as errors, thus degrading overall performance metrics. Enhanced annotation efforts to include comprehensive gold KG paths could potentially ameliorate this issue, leading to more accurate learning and prediction.

Table 7. Summary of major error types identified in CONVEX’s performance, highlighting the proportion of instances affected and the impact level on the results.

Error Type	Instances	Impact on Results
Semantically Similar Relations	35%	High
Absence of Gold Paths	25%	Critical

Table 7. Summary of major error types identified in CONVEX’s performance, highlighting the proportion of instances affected and the impact level on the results.

Error Type	Instances	Impact on Results
Semantically Similar Relations	35%	High
Absence of Gold Paths	25%	Critical

5.4.3. Improvement

Based on these findings, we propose several avenues for future research:

• Enhanced Relation Disambiguation: Developing more sophisticated techniques for relation extraction and disambiguation could help in more accurately distinguishing between similar relations. Incorporating additional contextual cues from the KG, such as entity types or relation aliases, may provide further discriminative power.
• Richer Training Data: Expanding the dataset to include more comprehensive gold KG paths could reduce the incidence of training and testing without adequate ground truth, thus improving the reliability of the learning process.
• Advanced Contextual Modeling: Leveraging newer models that better capture the nuances of conversational context may also help in more effectively discerning the relevance of different KG paths based on the dialogue history.

These strategies represent promising research directions that could substantially refine the efficacy of KG path ranking methodologies in conversational question-answering systems.

6. Conclusions and Future Work

Our research has critically addressed the utility of contrastive representation learning within the domain of conversational question answering over knowledge graphs. We posited the challenge as a KG path ranking problem and enriched the model’s context handling by incorporating comprehensive dialogue histories, fluent responses, and domain-specific information. The implemented model, named CONVEX, substantiates the premise that contrastive learning can markedly enhance performance in this intricate task.

Conclusions: The empirical results obtained from the CONVEX framework have demonstrably surpassed those of traditional baseline approaches, establishing a robust case for the efficacy of our method. The joint embedding of conversational context and KG paths within a unified representation space has been shown to significantly influence the outcome of ranking metrics positively. Moreover, detailed ablation studies have shed light on the individual contributions of various conversational elements, affirming that each component—complete dialog history, fluent responses, and domain information—plays a vital role in the model’s success.

Future Directions: Despite the advancements presented, our work unveils several avenues for future exploration and improvement: 1) Relation Extraction Enhancement: The error analysis highlighted challenges in accurately identifying and linking the correct KG relations. Future work could enhance relation extraction techniques by integrating richer KG contexts, akin to approaches seen in RECON [2]. Improving this aspect could refine the accuracy of the path ranking process. 2) Precision Enhancement: The disparity between the high Hit@5 scores and lower Precision@1 indicates potential improvements in the model’s precision. Innovating new methodologies or refining existing algorithms to better discriminate between closely ranked KG paths could help elevate the Precision@1 scores. 3) Exploring Additional Contextual Data: Integrating external knowledge sources or more expansive KG embeddings could provide deeper insights and more accurate context capturing, potentially leading to better model performance. 4) Multi-turn Dialogue Handling: Further research into handling multi-turn dialogues more effectively, perhaps by leveraging recent advancements in transformer models or recurrent neural strategies, could improve how context is maintained and utilized across longer conversations.

We believe that the insights garnered from the CONVEX framework will catalyze further research into this relatively underexplored area of IR within the ConvQA domain. By highlighting both the strengths and limitations of our approach, we hope to inspire future studies to build on our foundational work, pushing the boundaries of what is possible in conversational AI.