Multi-perspective Knowledge Extraction

• Multi-perspective Knowledge Extraction is a framework that decomposes tasks into distinct views—such as reasoning, explanation, and summarization—to augment sparse knowledge representations.
• It leverages large language models to generate and reintegrate perspective-specific outputs, enhancing knowledge graph completion, retrieval augmentation, and ICD coding accuracy.
• Empirical studies reveal substantial gains in recall and precision across legal, medical, and graph-based tasks, underscoring its practical impact and interpretability.

Multi-perspective Knowledge Extraction (MKE) denotes a class of methods that extract, generate, retrieve, or verify knowledge from multiple predefined perspectives and then reintegrate those perspective-specific signals into a downstream model. In recent arXiv work, the term appears most explicitly in description-based knowledge graph completion, where the task is to produce richer textual information from three perspectives—reasoning, explanation, and summarization—by querying a LLM and incorporating the outputs back into a knowledge graph; closely related formulations use multi-view retrieval in retrieval-augmented generation and multi-axial knowledge with evidence verification in ICD coding for Chinese electronic medical records (Xu et al., 2024, Chen et al., 2024, You et al., 19 Feb 2025).

1. Conceptual scope and recurring design pattern

Across these formulations, the central assumption is that a single representation is insufficient in knowledge-dense settings. In "Multi-perspective Improvement of Knowledge Graph Completion with LLMs" (Xu et al., 2024), the three perspectives are reasoning-based entity expansion, explanation-based relation understanding, and summarization-based structure extraction. In "Unlocking Multi-View Insights in Knowledge-Dense Retrieval-Augmented Generation" (Chen et al., 2024), the relevant construct is a set of predefined domain viewpoints or “professional perspectives,” such as “Medical History,” “Symptoms,” “Application of Law,” and “Focus of Dispute.” In "MKE-Coder: Multi-Axial Knowledge with Evidence Verification in ICD Coding for Chinese EMRs" (You et al., 19 Feb 2025), ICD codes are decomposed into four “axes”: Etiology axis, Anatomical site axis, Pathology axis, and Clinical manifestation axis.

These systems differ in target task, but they share a common workflow. First, a task-specific object is decomposed into multiple views, perspectives, or axes. Second, each perspective is used to produce an intermediate representation: rewritten queries in MVRAG, augmented entity or relation text and SameAs edges in MPIKGC, and axis-specific keyword sets plus evidence sentences in MKE-Coder. Third, the perspective-specific outputs are fused back into a conventional learner: a reader LLM in RAG, a description-based KGC scorer, or a prompt-tuned T5 verifier. This suggests that MKE is best understood as a family of augmentation-and-fusion strategies rather than a single canonical architecture.

Formulation	Perspective set	Reintegration target
MPIKGC	reasoning, explanation, summarization	description-based KGC model
MVRAG	predefined domain viewpoints	retrieval and final generation
MKE-Coder	four coding axes	evidence-supported ICD code inference

2. Formalization in description-based knowledge graph completion

In the knowledge-graph setting, the formal object is

$$G = \{ (h,r,t,d_h,d_r,d_t) \mid h,t\in\mathcal{E},\ r\in\mathcal{R},\ d_h,d_r,d_t\in\mathcal{D} \}.$$

A description-based KGC model learns an encoder $\Phi$ that maps textual inputs to embeddings,

$$e_h = \Phi(d_h \parallel name(h)), \quad v_r = \Phi(d_r \parallel name(r)), \quad e_t = \Phi(d_t \parallel name(t)),$$

together with a scoring function $$f : \mathbb{R}^d \times \mathbb{R}^d \times \mathbb{R}^d \to \mathbb{R}$$, and a typical margin-ranking loss

$$L = \sum_{(h,r,t)\in\mathcal{T}^+} \sum_{(h',r,t')\in\mathcal{T}^-} \max(0, \gamma + f(e_{h'},v_r,e_{t'}) - f(e_h,v_r,e_t)).$$

The MKE task is defined as: “Given a (possibly sparse) $G$ with minimal or missing $d_h, d_r, d_t$, produce richer textual information from three perspectives—reasoning (entity expansion), explanation (relation understanding), summarization (structure extraction)—by querying a LLM. Incorporate the LLM outputs back into $G$ so as to minimize $L$ and improve link-prediction and triplet-classification accuracy.” (Xu et al., 2024)

The reasoning module, MPIKGC-E, expands sparse entity descriptions through a Chain-of-Thought-style prompt: “Please provide all information about {Entity Name}. Give the rationale before answering:”. The generated paragraph $d_h^{LLM}$ is concatenated to the original $\Phi$0 or the entity name alone before encoding via $\Phi$1. The explanation module, MPIKGC-R, augments relation names with three one-sentence explanations: a global prompt, a local triplet rephrasing prompt, and a reverse-direction prompt. The integration step is

$\Phi$2

The summarization module, MPIKGC-S, extracts five representative keywords from an entity description, computes

$\Phi$3

selects the top-$\Phi$4 highest-scoring pairs, and adds triples $\Phi$5 together with self-loops $\Phi$6 to $\Phi$7. After augmentation, the loss remains a margin-ranking or contrastive loss over $\Phi$8.

This formulation is notable for being model-agnostic at the KGC layer. The backbones evaluated include KG-BERT, SimKGC, LMKE, and CSProm-KG, while the LLMs used for generation include LLaMA-2-7B-chat, ChatGLM2-6B, ChatGPT (gpt-3.5-turbo), and GPT-4. The perspective modules therefore function as textual and structural augmentation components rather than replacements for the scoring model.

3. Multi-view retrieval as perspective-conditioned knowledge extraction

In retrieval-augmented generation, the multi-perspective extraction problem is framed as intention-aware query rewriting and weighted retrieval. MVRAG extends a standard RAG pipeline by inserting two additional “multi-view” stages—Intention Recognition and Query Rewriting—before retrieval, and a weighted re-ranking step after retrieval. The pipeline is: input query $\Phi$9 to Intention Recognition to produce a Perspective Vector $$e_h = \Phi(d_h \parallel name(h)), \quad v_r = \Phi(d_r \parallel name(r)), \quad e_t = \Phi(d_t \parallel name(t)),$$0; for each non-zero entry in $$e_h = \Phi(d_h \parallel name(h)), \quad v_r = \Phi(d_r \parallel name(r)), \quad e_t = \Phi(d_t \parallel name(t)),$$1, produce a subquery $$e_h = \Phi(d_h \parallel name(h)), \quad v_r = \Phi(d_r \parallel name(r)), \quad e_t = \Phi(d_t \parallel name(t)),$$2; retrieve a document set $$e_h = \Phi(d_h \parallel name(h)), \quad v_r = \Phi(d_r \parallel name(r)), \quad e_t = \Phi(d_t \parallel name(t)),$$3 per subquery; re-rank all retrieved documents using both similarity and perspective weights; and concatenate the top documents with the original query into a prompt $$e_h = \Phi(d_h \parallel name(h)), \quad v_r = \Phi(d_r \parallel name(r)), \quad e_t = \Phi(d_t \parallel name(t)),$$4 for the reader LLM (Chen et al., 2024).

The perspective vector is

$$e_h = \Phi(d_h \parallel name(h)), \quad v_r = \Phi(d_r \parallel name(r)), \quad e_t = \Phi(d_t \parallel name(t)),$$5

with

$$e_h = \Phi(d_h \parallel name(h)), \quad v_r = \Phi(d_r \parallel name(r)), \quad e_t = \Phi(d_t \parallel name(t)),$$6

Here $$e_h = \Phi(d_h \parallel name(h)), \quad v_r = \Phi(d_r \parallel name(r)), \quad e_t = \Phi(d_t \parallel name(t)),$$7 is the set of predefined domain viewpoints, $$e_h = \Phi(d_h \parallel name(h)), \quad v_r = \Phi(d_r \parallel name(r)), \quad e_t = \Phi(d_t \parallel name(t)),$$8 is a relevance score, and $$e_h = \Phi(d_h \parallel name(h)), \quad v_r = \Phi(d_r \parallel name(r)), \quad e_t = \Phi(d_t \parallel name(t)),$$9 is a small threshold such as $$f : \mathbb{R}^d \times \mathbb{R}^d \times \mathbb{R}^d \to \mathbb{R}$$0. For each $$f : \mathbb{R}^d \times \mathbb{R}^d \times \mathbb{R}^d \to \mathbb{R}$$1 with $$f : \mathbb{R}^d \times \mathbb{R}^d \times \mathbb{R}^d \to \mathbb{R}$$2, the viewpoint-aware subquery is

$$f : \mathbb{R}^d \times \mathbb{R}^d \times \mathbb{R}^d \to \mathbb{R}$$3

For each non-zero perspective weight $$f : \mathbb{R}^d \times \mathbb{R}^d \times \mathbb{R}^d \to \mathbb{R}$$4, the retriever returns

$$f : \mathbb{R}^d \times \mathbb{R}^d \times \mathbb{R}^d \to \mathbb{R}$$5

with $$f : \mathbb{R}^d \times \mathbb{R}^d \times \mathbb{R}^d \to \mathbb{R}$$6, for example $$f : \mathbb{R}^d \times \mathbb{R}^d \times \mathbb{R}^d \to \mathbb{R}$$7. Each document is then re-scored as

$$f : \mathbb{R}^d \times \mathbb{R}^d \times \mathbb{R}^d \to \mathbb{R}$$8

and the top $$f : \mathbb{R}^d \times \mathbb{R}^d \times \mathbb{R}^d \to \mathbb{R}$$9 pooled documents are concatenated to form

$$L = \sum_{(h,r,t)\in\mathcal{T}^+} \sum_{(h',r,t')\in\mathcal{T}^-} \max(0, \gamma + f(e_{h'},v_r,e_{t'}) - f(e_h,v_r,e_t)).$$0

A distinctive feature of this formulation is that no additional learned embeddings are required: the viewpoints $$L = \sum_{(h,r,t)\in\mathcal{T}^+} \sum_{(h',r,t')\in\mathcal{T}^-} \max(0, \gamma + f(e_{h'},v_r,e_{t'}) - f(e_h,v_r,e_t)).$$1 enter as natural-language descriptors, and MVRAG uses two LLM calls per perspective, one for scoring and one for rewriting. The paper notes that one could optionally embed $$L = \sum_{(h,r,t)\in\mathcal{T}^+} \sum_{(h',r,t')\in\mathcal{T}^-} \max(0, \gamma + f(e_{h'},v_r,e_{t'}) - f(e_h,v_r,e_t)).$$2 and learn a small adapter, but the framework treats LLMs as black boxes. In knowledge-dense domains such as law and medicine, this design makes the perspective labels directly inspectable and links retrieved documents to explicit professional viewpoints.

4. Multi-axial extraction and evidence verification for ICD coding

MKE-Coder applies the MKE idea to automatic ICD coding for Chinese EMRs by decomposing each candidate code $$L = \sum_{(h,r,t)\in\mathcal{T}^+} \sum_{(h',r,t')\in\mathcal{T}^-} \max(0, \gamma + f(e_{h'},v_r,e_{t'}) - f(e_h,v_r,e_t)).$$3 into four mappings: Etiology axis $$L = \sum_{(h,r,t)\in\mathcal{T}^+} \sum_{(h',r,t')\in\mathcal{T}^-} \max(0, \gamma + f(e_{h'},v_r,e_{t'}) - f(e_h,v_r,e_t)).$$4, Anatomical site axis $$L = \sum_{(h,r,t)\in\mathcal{T}^+} \sum_{(h',r,t')\in\mathcal{T}^-} \max(0, \gamma + f(e_{h'},v_r,e_{t'}) - f(e_h,v_r,e_t)).$$5, Pathology axis $$L = \sum_{(h,r,t)\in\mathcal{T}^+} \sum_{(h',r,t')\in\mathcal{T}^-} \max(0, \gamma + f(e_{h'},v_r,e_{t'}) - f(e_h,v_r,e_t)).$$6, and Clinical manifestation axis $$L = \sum_{(h,r,t)\in\mathcal{T}^+} \sum_{(h',r,t')\in\mathcal{T}^-} \max(0, \gamma + f(e_{h'},v_r,e_{t'}) - f(e_h,v_r,e_t)).$$7. The multi-axis knowledge is represented as

$$L = \sum_{(h,r,t)\in\mathcal{T}^+} \sum_{(h',r,t')\in\mathcal{T}^-} \max(0, \gamma + f(e_{h'},v_r,e_{t'}) - f(e_h,v_r,e_t)).$$8

where each $$L = \sum_{(h,r,t)\in\mathcal{T}^+} \sum_{(h',r,t')\in\mathcal{T}^-} \max(0, \gamma + f(e_{h'},v_r,e_{t'}) - f(e_h,v_r,e_t)).$$9 is the small vocabulary of keywords or phrases defining axis $G$0. In practice, these are obtained by splitting the code string according to the official ICD indexing rules, looking up each segment in an ICD-code→description dictionary, extracting noun phrases or headwords, and expanding each keyword set with clinical synonyms via UMLS or manual synonym lists. For example, code H31.403 is decomposed into Etiology $G$1“inflammatory,” “degenerative”$G$2, Anatomical site $G$3“retina,” “choroid”$G$4, Pathology $G$5“morphological change,” “atrophy”$G$6, and Clinical manifestation $G$7“vision loss,” “floaters”$G$8 (You et al., 19 Feb 2025).

The end-to-end system begins with preprocessing: Chinese word segmentation with jieba, removal of rare symbols, normalization of digits and units, sentence splitting on punctuation marks, and section tagging by source field. Candidate code pre-selection uses

$G$9

with

$d_h, d_r, d_t$0

where $d_h, d_r, d_t$1, $d_h, d_r, d_t$2, and $d_h, d_r, d_t$3. For evidence retrieval, the overall axis keyword set is

$d_h, d_r, d_t$4

Within each likely section $d_h, d_r, d_t$5, sentences are scored by

$d_h, d_r, d_t$6

and the best sentence is

$d_h, d_r, d_t$7

Sentences are greedily accumulated into $d_h, d_r, d_t$8 until every keyword in $d_h, d_r, d_t$9 has appeared or the section list is exhausted.

Evidence filtering uses a Chinese SimBERT model fine-tuned in two stages. At inference time,

$G$0

and

$G$1

The final evidence set is

$G$2

with threshold $G$3. Code verification is then converted into a fill-in-the-blank problem using a T5 encoder and soft prompt tokens:

$G$4

The verbalizer is $G$5“no”, “yes”$G$6, the prediction is

$G$7

and the objective is

$G$8

Only the soft tokens are learned; all T5 weights remain frozen.

5. Empirical findings and ablation evidence

In multi-view RAG, the reported gains are large on both legal and medical case retrieval. On LeCaRDv2, bge-m3 rises from 3.13% Recall@100, 0.94% Precision@100, and 1.44% F1 to 16.53%, 4.96%, and 7.63% with MVRAG. On PMC-Patients, gte-large rises from 1.60% Recall@100, 0.11% Precision@100, and 0.20% F1 to 13.05%, 2.13%, and 3.67%. The paper states that all models saw consistent, large relative gains, often 4–10× improvements in Recall@100. An ablation that removes one perspective at a time shows that omitting “Fundamental Facts” in the legal task drops Recall@100 from 16.53% to 12.11% (Chen et al., 2024).

In description-based KGC, the gains are smaller in absolute magnitude but consistent across backbones and tasks. On FB15k-237 with CSProm-KG, the original model reports MR 188, MRR 35.23, H@1 26.05, H@3 38.72, and H@10 53.57; +MPIKGC-E gives MR 195, MRR 35.51, H@1 26.38, H@3 38.96, and H@10 53.74; +MPIKGC-R gives MR 192, MRR 35.38, H@1 26.29, H@3 38.83, and H@10 53.50; +MPIKGC-S gives MR 179, MRR 35.95, H@1 26.71, H@3 39.52, and H@10 54.30. On triplet classification with KG-BERT on FB13, the original accuracy is 84.74%, +MPIKGC-E is 86.29%, +MPIKGC-R is 84.51%, and +MPIKGC-S is 85.35%. The LMKE ablation on FB15k-237 shows that +S gives the largest single-perspective net gain on Hits@10, while the pairwise combination R+S achieves the best MRR and H@1 (Xu et al., 2024).

In ICD coding, MKE-Coder is evaluated on a 70 K-record Chinese EMR set with Macro Precision, Recall, F1, Accuracy, and Precision@5. The reported state-of-the-art results are 72.61% macro-F1, 74.32% accuracy, and 92.73% P@5. Interpretability is quantified by expert review, which reports 50% fully supported and 46% partially supported evidence. The practical evaluation within simulated real coding scenarios states that the method significantly aids coders in enhancing both their coding accuracy and speed (You et al., 19 Feb 2025).

6. Interpretability, limitations, and points of clarification

A central motivation for MKE is interpretability. In MVRAG, multi-view retrieval “surfaces” which domain perspectives contributed to each retrieved document via the weight $G$9, making it clear, for instance, that a case was retrieved chiefly because it matched “Symptoms” rather than “Laboratory Data.” The Huntington’s disease case study illustrates the point: vanilla RAG retrieved Vitamin B12–deficiency articles because of superficial symptom match, whereas MVRAG’s symptoms and history perspectives directed retrieval toward neurodegeneration and genetic-testing literature, avoiding misdiagnosis (Chen et al., 2024).

In MPIKGC, interpretability is distributed across prompts and augmentation channels rather than a single explanation module. Entity expansion yields explicit fact-rich paragraphs; relation understanding provides global, local, and reverse-direction semantics; and structure extraction adds SameAs edges through a transparent keyword-overlap rule. A possible source of confusion is the status of those SameAs edges. Within MPIKGC-S, they are generated by selecting top-$L$0 entity pairs according to keyword-set similarity and then treating the resulting triples as additional positive facts appended to the training set. The example pairing of “Michael Bay” and “Transformers: Dark of the Moon” follows this scoring procedure, not a separate entity-resolution algorithm (Xu et al., 2024).

In MKE-Coder, interpretability is tied to evidence verification rather than only to candidate ranking. Each supported code is returned with its evidence set $L$1, where evidence is first retrieved by axis-keyword coverage and then filtered by Clinical-SimBERT. The final T5 module verifies whether all the axis knowledge associated with the candidate code is supported by evidence and provides recommendations accordingly. This suggests a stricter notion of support than retrieval-only systems, because the framework explicitly asks whether a candidate code is supported by evidence rather than merely similar to the diagnosis text (You et al., 19 Feb 2025).

Taken together, these works indicate that multi-perspective extraction is most useful when sparse inputs, ambiguous labels, or heterogeneous domain evidence make one-view modeling unreliable. They also indicate that perspective choice is task-dependent: professional viewpoints for RAG, reasoning/explanation/summarization for KGC, and four coding axes for ICD inference. The current literature therefore treats “perspective” less as a fixed ontology than as a structured decomposition chosen to expose otherwise missing or weakly represented knowledge.