Personal Textual Knowledge Bases

• Personal Textual Knowledge Bases (PTKBs) are user-centered collections of natural-language statements that capture personal preferences, constraints, and traits.
• They support personalized services by enabling tailored conversational information retrieval, query reformulation, and long-term knowledge management.
• PTKB implementations range from manual note-taking and automated extraction to graph-based representations and iterative editing for improved retrieval performance.

Searching arXiv for PTKB-related papers and the cited work to ground the encyclopedia entry. Personal Textual Knowledge Bases (PTKBs) are user-centered knowledge resources in which personal knowledge is expressed or derived through text and then used to support personalized services, conversational assistants, retrieval, and long-term knowledge management. In the recent literature, the term appears most explicitly in personalized conversational information retrieval, where a PTKB is a finite set of natural-language statements about a user; closely related lines of work treat personal note collections as “second brains,” represent research-centered personal knowledge as graphs extracted from textual sources, or combine symbolic triples with longer textual descriptions and aspect-specific passages in a personalized knowledge base (Aliannejadi et al., 2024, Mo et al., 2024, Ferreira et al., 24 Sep 2025, Chakraborty et al., 2022, Wang et al., 2023, Skjæveland et al., 2023). Across these variants, the central problem is consistent: how to externalize, organize, retrieve, update, and govern personally relevant knowledge encoded in or linked to text.

1. Definition and conceptual boundaries

In the TREC iKAT setting, a PTKB—also called a “Personalized Topic Knowledge Base”—is a set of narrative sentences that encode high-level personal information about a user or persona relevant to a decisional search task. The sentences are assumed to be collected from previous conversations with the user and may encode preferences, constraints, goals, or attitudes, but they deliberately exclude personally identifiable information. At the formal level used in later personalized CIR work, the PTKB is written as

$\mathcal{U} = \{s_t\}_{t=1}^T,$

where each $s_t$ is a natural-language sentence describing characteristics or preferences of the user; the iKAT design also supports the equivalent persona-indexed view

$$\text{PTKB}_p = \{ s_{p,1}, s_{p,2}, \dots, s_{p,n_p} \}.$$

These formulations emphasize that a PTKB is sentence-level, textual, and unstructured relative to a knowledge graph schema (Aliannejadi et al., 2024, Mo et al., 2024).

This sentence-based view is narrower than the broader PKG literature but not incompatible with it. A personal knowledge graph is defined as a knowledge graph where a single individual has full read and write access and the exclusive right to grant others read and write access to specified parts, with the primary purpose of delivering services customized to that owner. That definition is intentionally content-agnostic: the graph is “personal” because of ownership and purpose, not because every fact must be attached to a central user node. This distinguishes a PKG from a personalized knowledge graph owned by a service and derived as a user-specific subgraph of a larger public or proprietary KG (Skjæveland et al., 2023).

Two recurrent misconceptions are explicitly rejected by the literature. First, PTKBs are not synonymous with structured KGs: in iKAT they are textual and narrative rather than triples. Second, personalization is not uniformly useful: not every conversational turn requires the PTKB, and not every query benefits from injecting all personal knowledge into retrieval or generation (Aliannejadi et al., 2024, Mo et al., 2024).

2. Representational forms

Recent work uses several representational forms for PTKB-like systems.

Form	Representation	Exemplars
Sentence-level persona PTKB	Natural-language statement set $\mathcal{U}=\{s_t\}_{t=1}^T$	iKAT, personalized CIR
Note-centered PKB / “second brain”	Notes organized with folders, tags, links, templates, daily notes	Obsidian case study
Graph-based personal KB	$$\mathcal{G}=\{\mathcal{E}, \mathcal{R}, \mathcal{F}\}$$ with entities, relations, facts	Personal Research Knowledge Graph
Hybrid personalized KB	Entity descriptions, relational triples, entity-aspect content	KnowledGPT

The sentence-level form is deliberately minimal. In iKAT, PTKB statements are high-level, task-relevant, and privacy-preserving. Their granularity is roughly one sentence, and there is no fine-grained slot structure or labeled attributes; systems must interpret semantics directly from text. This makes PTKB entries easy to place into prompts or query rewrites but leaves relation typing implicit (Aliannejadi et al., 2024, Mo et al., 2024).

The note-centered form appears in studies of personal knowledge bases and “second brains.” In an Obsidian-based case study with seven computer science researchers, participants managed vaults containing folders, tags, links, templates, daily notes, and sometimes multiple vaults. The paper defines a PKB as “the collection of records for future reference,” an ePKB as electronic tools for expressing, capturing, and retrieving such knowledge, and a “second brain” as a metaphor emphasizing cognitive extension. A central empirical finding is that participants’ retrieval strategy influences how they build and maintain content: folder-centric users maintain hierarchies, tag-centric users invest in metadata schemes, search-centric users rely on descriptive note text, and link-centric users create associative trails (Ferreira et al., 24 Sep 2025).

The graph-based form specializes personal knowledge to explicit entities and relations. In the Personal Research Knowledge Graph framework, a PRKG is a user-centric KG

$$\mathcal{G} = \{\mathcal{E}, \mathcal{R}, \mathcal{F}\},$$

centered on a researcher node and constrained to research-related activities, resources, and fine-grained research knowledge. The paper highlights a recurring research tuple pattern, $(task, method, tool, dataset, metric)$, and recommends temporalized relations $r:[t_1,t_2]$ for jobs, projects, and changing interests (Chakraborty et al., 2022).

KnowledGPT adopts a hybrid symbolic-textual representation. Its personalized KB stores three kinds of material: entity descriptions, relational triples $(h,r,t)$, and entity-aspect information $(e,a,c)$, where the object of the aspect record is long text. On a HotpotQA-derived extraction study, triples alone captured 0.53 recall, adding entity descriptions raised recall to 0.66 for ChatGPT and 0.62 for GPT-4, and adding entity-aspect information increased recall to 0.81 and 0.86 respectively. The result is a PTKB representation in which structured and longer textual views coexist (Wang et al., 2023).

3. Population and maintenance

PTKB construction ranges from manual authoring to fully automated extraction. In iKAT, PTKBs are hand-crafted by organizers as part of topic creation: a decisional topic is selected, personas are designed, a PTKB is created as narrative sentences capturing high-level traits and constraints, and then user turns, relevant PTKB statements, retrieved passages, and canonical responses are constructed around that PTKB (Aliannejadi et al., 2024).

Several works treat conversational text as a primary population source. DSCGN predicts personal attributes such as profession and hobby from user utterances without labeled utterances by combining document-level distant supervision from Wikipedia with contextualized word-level supervision derived by a label guessing method; the output is a ranked list of attribute values that can populate subject-predicate-object triples in a personal KB. On the profession dataset with Wiki-category supervision, DSCGN reaches MRR 0.44 and nDCG 0.60, compared with 0.35 and 0.55 for the best supervised CHARM$s_t$0 baseline reported there (Liu et al., 2022). PEARL addresses the same task in a stricter low-resource setting with no labeled utterances and no external data, using biterm semantic acquisition and an iterative attribute-knowledge-integrated biterm topic model; on the profession task it reports MRR 0.49 and nDCG 0.64, and on hobby MRR 0.31 and nDCG 0.54 (Liu et al., 2022).

Cold-start personal KB population from unstructured text has also been treated as a hybrid IE and graph-completion problem. A TACRED-based system combines SystemT rule-based annotators, a fine-grained personal-data entity classifier, graph construction, and GNN link prediction. On fine-grained entity typing, augmenting the classifier with token-level features and SystemT types raises Macro F1 from 0.678 to 0.740 and Micro F1 from 0.617 to 0.672. For link prediction, a position-aware GNN reaches ROC AUC 0.6473 versus 0.4047 for a GCN on the augmented graph. In cold-start slot-filling evaluation, the augmented-plus-GNN system reaches Hop-all precision 1.00, recall 0.657, and F1 0.791; for protected attributes, recall rises from 0.095 on baseline TACRED to 0.972 with precision 1.00 (Vannur et al., 2020).

PRKG work emphasizes multi-source extraction from CVs, local devices, conversations with a chatbot, downloads, search queries, emails, meeting transcripts, and papers. A concrete implemented component uses SpERT trained on SciERC to extract scientific entities and relations from papers read by the researcher; low-confidence extractions are intended to trigger active confirmation from the user before insertion (Chakraborty et al., 2022). By contrast, second-brain studies stress progressive organization rather than one-shot extraction: users capture quickly, then later process, tag, link, summarize, or discard notes (Ferreira et al., 24 Sep 2025).

Maintenance is increasingly treated as a first-class problem. STACKFEED models a textual KB as a set of documents $s_t$1, each chunked as $s_t$2, and edits the KB rather than the base LLM. Its actor–critic scheme assigns one actor per document and allows only structured actions—EditChunk, AddChunk, DeleteChunk—while a centralized critic converts failure feedback into document-specific textual gradients. On several KB-editing benchmarks, STACKFEED improves test accuracy over baselines and improves KB coherence, for example reaching 37.04 versus 32.22 test accuracy on Pony and 59.38 versus 53.40 on SciPy relative to PromptAgent-E (Gupta et al., 2024). This suggests that PTKB maintenance can be formalized as iterative editing with explicit diffs, provenance, and feedback loops.

4. Retrieval, reformulation, and interaction

In personalized conversational IR, the PTKB is used together with conversation history to reformulate the current turn. The formal problem is: given current query $s_t$3, history

$s_t$4

and PTKB

$s_t$5

produce a personalized stand-alone reformulated query $s_t$6 for retrieval (Mo et al., 2024).

A central finding is that PTKB is noisy: not every sentence is relevant to the current query. The literature therefore studies PTKB sentence selection before query rewriting. In one comparison, “None” ignores PTKB, “Use all” concatenates the whole PTKB, “Human” uses human-labeled relevant sentences, “Automatic” defines relevance by positive retrieval impact, and STR/SAR are LLM-based strategies that select and reformulate either in two stages or jointly. On the whole iKAT test set with BM25, “None” reaches MRR 44.35, NDCG@3 21.22, NDCG@5 20.68, and MAP 8.91, while naïve “Use all” falls to 40.36, 19.19, 18.84, and 8.28. On the 67-turn “need PTKB” subset, however, Automatic reaches MRR 33.72, NDCG@3 16.78, NDCG@5 16.28, and MAP 5.33, outperforming the non-personalized baseline. With high-quality in-context guidance, SAR 3-shot reaches MRR 45.73, NDCG@3 22.72, NDCG@5 22.01, and MAP 10.37 on BM25 over the whole test set, exceeding the best no-PTKB baseline (Mo et al., 2024).

The iKAT collection operationalizes PTKB use at three task levels: PTKB statement ranking, passage retrieval or ranking, and response generation. Systems are expected to retrieve persona memory, rewrite the query with PTKB and dialogue context, retrieve external passages from ClueWeb22, and generate grounded, persona-appropriate responses. The task targets decisional search, where relevance is conditional on persona. A motivating pattern is that “vegan,” “diabetic,” or “environmentalist” yields different relevant options for the same question about milk alternatives (Aliannejadi et al., 2024).

KnowledGPT extends the operational picture by turning LLM access to a personalized KB into program generation. The controller produces a Python search() program using three abstract functions—get_entity_info, find_entity_or_value, and find_relationship—and can also store knowledge extracted from text into a personalized KB. For sparse retrieval it uses rewrite plus response; for dense retrieval it uses the rewrite alone. The architecture is explicitly designed to support both retrieval from and storage into a personalized KB (Wang et al., 2023).

In note-based PTKBs, interaction is often simpler but structurally related. The Obsidian study identifies four retrieval methods—search box, tag list or tag-based search, in-note tag navigation, and in-note link navigation—and reports that users combine them with hierarchical browsing. The core empirical claim is that retrieval strategy shapes construction, so representation and access cannot be treated as separate design layers (Ferreira et al., 24 Sep 2025).

5. Evaluation and benchmarking

The most explicit PTKB benchmark is TREC iKAT 2023. The collection contains 36 personalized dialogues over 20 topics, with 1–3 personas per topic and effectively 36 PTKBs, each tied to one topic–persona pair and one dialogue. It contains 427 total turns, of which 326 are considered for passage relevance and 176 are judged by NIST for passages; average dialogue length is 13.04 turns. The underlying ClueWeb22-B subset contains 116,838,987 passages, and the judged passage pool contains 26,159 passages (Aliannejadi et al., 2024).

PTKB relevance is assessed at statement level. Organizers annotate every turn with relevant PTKB statements, and NIST judges PTKB relevance for turns selected for passage relevance. Only a subset of turns are PTKB-dependent; these are explicitly marked as personalized turns. PTKB statement ranking is evaluated with $s_t$7, $s_t$8, $s_t$9, and $$\text{PTKB}_p = \{ s_{p,1}, s_{p,2}, \dots, s_{p,n_p} \}.$$0. Passage ranking uses $$\text{PTKB}_p = \{ s_{p,1}, s_{p,2}, \dots, s_{p,n_p} \}.$$1, $$\text{PTKB}_p = \{ s_{p,1}, s_{p,2}, \dots, s_{p,n_p} \}.$$2, full-depth nDCG, $$\text{PTKB}_p = \{ s_{p,1}, s_{p,2}, \dots, s_{p,n_p} \}.$$3, $$\text{PTKB}_p = \{ s_{p,1}, s_{p,2}, \dots, s_{p,n_p} \}.$$4, Recall, and mAP, with mean nDCG@5 as the primary metric. Response generation is assessed on four dimensions: human-annotated relevance and completeness, GPT-4-annotated naturalness, and groundedness with respect to provenance passages (Aliannejadi et al., 2024).

The benchmark also exposes a specific methodological controversy: intuitive human relevance labels for PTKB sentences need not align with retrieval effectiveness. In the personalized CIR study, overlap between human, automatic, and LLM labels is low, and Automatic relevance—defined by whether adding a PTKB sentence improves retrieval metrics—often outperforms human selection when the query genuinely needs personalization. This does not invalidate human judgment; rather, it shows that “relevant to the user profile” and “helpful for retrieval” are not identical operational categories (Mo et al., 2024).

Evaluation work outside iKAT also points to broader PTKB quality axes. STACKFEED evaluates KB editing by completeness, generalization, and coherence, using train accuracy, held-out test accuracy, and document-wise G-Eval coherence scores respectively (Gupta et al., 2024). Extensive materialization of model knowledge, though not a PTKB paper, analyzes a generated KB along scale, accuracy, bias, temporal cutoff, and consistency, including inverse-relation consistency; this suggests comparable auditing dimensions for future PTKBs, especially when they are maintained automatically (Hu et al., 2024).

6. Governance, privacy, and open problems

Privacy and control are foundational rather than ancillary. In iKAT, PTKBs are restricted to high-level personal information and explicitly exclude personally identifiable information (Aliannejadi et al., 2024). The broader PKG literature goes further: a PKG is defined by single-person ownership, full read-write access by that owner, and the owner’s exclusive right to grant access to specified parts for others, with personalized service delivery as the primary purpose (Skjæveland et al., 2023). This makes PTKB governance a question of technical architecture as much as of representation.

Role-based sharing has been demonstrated in graph-based personal knowledge systems. The PRKG prototype implements Neo4j role-based access control so that a collaborator can be denied access to selection committee nodes, reviewer-of edges, and all write operations; the paper also notes that Neo4j lacks property-level RBAC, leaving finer-grained exposure control as future work (Chakraborty et al., 2022). In note-based settings, the governance problem appears in another form: users confront blank-page overhead, dynamic reorganization, fragmentation across multiple vaults, and the need to refactor structure as mental models evolve (Ferreira et al., 24 Sep 2025).

Several open problems recur across the literature. PTKBs in iKAT are small, static, and topic-scoped rather than long-term evolving memories (Aliannejadi et al., 2024). Personalized CIR shows that systems still struggle to decide when personalization is needed at all, how to select retrieval-effective PTKB sentences, and how to avoid degrading performance by blindly injecting the whole PTKB (Mo et al., 2024). Population methods from conversations and enterprise text raise continuing concerns about privacy, consent, fairness, and user burden (Liu et al., 2022, Liu et al., 2022, Vannur et al., 2020). Second-brain studies indicate that organization strategy and retrieval strategy must be co-designed, not optimized in isolation (Ferreira et al., 24 Sep 2025). A plausible implication is that future PTKB systems will need to combine sentence-level persona memory, note-centric long-form material, graph-based normalization, explicit provenance, and feedback-driven editing within a governance model where the owner can inspect, correct, and selectively expose both text and derived knowledge.