Financial filings report the same quantity at several time resolutions side by side. A quarterly report (10-Q) shows "three months ended September 30" next to "nine months ended September 30," and the annual report (10-K) adds the full fiscal year. Questions rarely use the filing's wording. A user asks about "Q3 revenue," "third-quarter sales," or "revenue in the quarter after the acquisition closed." Retrieval systems handle this in two ways. Some match words, which fails when "Q3" never appears in the table. Others rank evidence by how central it is in a graph of linked facts, which creates a quieter failure: coarse aggregates such as year-to-date and annual totals connect to many finer facts, so they rise to the top of the ranking. The model then answers a quarterly question with a nine-month number. This note proposes to measure that failure and to fix it with a ranking that is conditioned on the time resolution the question actually needs, combined with a canonicalizer that maps period phrases to explicit date intervals.
In two sentences: graph-centrality retrieval systematically favors low-frequency, aggregated facts, and in multi-hop financial QA this turns the vocabulary gap around time periods into wrong-but-plausible answers. We propose a personalized random-walk ranking that only spreads importance among facts at the resolution the question requires, with that resolution resolved hop by hop. We also test whether a simple period filter makes the graph unnecessary.
Core Claim
Recent work shows that centrality is useful scaffolding for language models. A Systematic Exploration of Knowledge Graph Alignment with Large Language Models in Retrieval Augmented Generation analyzed 81 graph features across 15 LLMs. It found that high-degree and PageRank nodes act as "reasoning anchors," and that organizing linearized graphs around them gives a 7.3% average accuracy gain. Its graphs are static and carry no notion of time resolution. Our first claim, which motivates everything else, is a hypothesis to be measured: when a fact graph mixes resolutions, a node summarizing a longer period links to every finer node it contains. Those nodes become the anchors, which biases evidence toward coarse aggregates. This is the graph version of the low-pass-filter problem described in Direction 1: aggregation hides short-horizon effects. Decision support from financial disclosures with deep neural networks and transfer learning and The "two sessions": institutional investors selloff to avoid ambiguity document this in forecasting and ownership data.
Our second claim is that the fix needs two parts working together. The first part is lexical. Period phrases must be canonicalized into intervals with a resolution label, because the vocabulary gap between question and filing is largely a gap in how time is named. The second part is structural, and it is the part a skeptic will doubt. The ranking must restrict where importance can flow, according to a target resolution. That target is often unknown until an intermediate hop is resolved (for example, "the quarter after the acquisition"). For this reason a filter applied once at query time is not enough.
Research Program
Graph construction. We build a fact graph from the filings of each company. Nodes are (entity, line item, interval) facts taken from table cells, together with the text spans that mention them. Edges connect facts that share an entity or line item, facts linked by an explicit textual reference, and facts linked by interval containment ("Q3 is contained in the nine-month period, which is contained in the fiscal year"). Containment edges are typed as aggregation edges. Line-item synonyms ("net sales," "revenue," "total revenues") are merged with an entity-fusion step, in the spirit of EventRAG. Unlike Hierarchical Lexical Graph for Enhanced Multi-Hop Retrieval, we do not decompose documents into propositions with an LLM. We reuse the table structure the filings already provide, which keeps construction cheap.
Period canonicalizer. A rule-based parser handles the regular phrasings ("three months ended," "fiscal 2020," "year-to-date") and uses the company's fiscal calendar. A small LLM fallback handles relative expressions. Each phrase is mapped to an interval and a resolution $g \in$ {quarter, year-to-date, year}. This is where the vocabulary-mismatch direction enters the work. The canonicalizer targets the implicit, non-overlapping phrasing that FinQA and TAT-QA identify as a major source of retrieval and evidence-extraction errors. TAT-QA attributes 55% of failures to evidence extraction, and FinQA attributes about 15% to the retriever.
Ranking. At each hop we score nodes with a personalized PageRank:
$$r = (1-\alpha)\,P_{g}^{\top} r + \alpha\, s_{g}$$
The symbols are defined as follows. $r$ is the vector of node scores. $\alpha$ is the restart probability. $s_g$ is a seed distribution placed on question entities whose interval has the target resolution $g$. $P_g$ is the transition matrix after each edge $(u,v)$ is reweighted by a compatibility factor $w(g_v, g)$. That factor is 1 when the resolutions match. It is a small constant $\epsilon$ on aggregation edges that lead to a different resolution, so that importance can reach a coarse fact only when the question needs it. When $\epsilon = 1$, the method reduces to the static centrality anchoring of Tian et al., and this is the ablation that isolates our contribution.
Hop-by-hop resolution. For multi-hop questions, a reader model proposes the next sub-question after each hop, in the way IRCoT (Interleaving Retrieval with Chain-of-Thought Reasoning for Knowledge-Intensive Multi-Step Questions) interleaves reasoning and retrieval. The canonicalizer then turns any interval the previous hop produced (for example, an acquisition date) into the next $g$ and the next seed set. The top-ranked subgraph is linearized around its highest-scoring anchors, using the template ranking reported by Tian et al. That paper found a 10.7-point spread between the best and worst templates, so we fix the template across all conditions.
Baselines
Naive reference. The do-nothing reference is dense chunk retrieval followed by an LLM reader. MultiHop-RAG shows this setup tops out at Hits@10 of 0.7467 and 0.56 GPT-4 accuracy on news-domain multi-hop queries. We run it with the same embedding model, the same reader, and the same top-k as every other condition.
Strongest structured-retrieval competitor. On filings, this is StatementGraphRAG from Hierarchical Lexical Graph for Enhanced Multi-Hop Retrieval. It runs beam search over entity-linked propositions and was evaluated on SEC-10Q among five datasets, reaching 73.6% average correctness across those five. No reported result was found on Lacuna for SEC-10Q alone. We re-implement it on our corpus with the same reader and context budget.
Strongest heterogeneous table-and-text competitor. This is RELOOP. It navigates a hierarchical sequence encoding of tables and text with a planner and a sufficiency check, and reports 75.7% accuracy and 83.5 F1 on TAT-QA. Beating it means exceeding 83.5 F1 on TAT-QA with a comparable number of retrieval steps. RELOOP averages about 4 steps on HotpotQA, and we will report steps for both systems.
Agentic competitor. This is an AT2QA-style loop (Let the Agent Search). The agent gets a search tool that supports symbolic period filtering over our canonicalized intervals. AT2QA reached 88.7% Hits@1 on MultiTQ, but over clean timestamped quadruples rather than filings. We give it the same reader model and a matched token budget, and we report accuracy at equal LLM-call counts.
Isolating ablations. Static centrality anchoring ($\epsilon = 1$, canonicalizer on) isolates the resolution-conditioned walk. Canonicalizer plus flat dense retrieval with a hard period filter, and no graph, isolates the structural contribution. For older reference points, we cite FinQANet at 61.24% execution accuracy on FinQA and TAGOP at 58.0 F1 on TAT-QA. All tuning (choices of $\alpha$, $\epsilon$, and k) uses a held-out development split. Every baseline gets the same number of tuning trials.
Why not just…?
Why not just filter by normalized period metadata and skip the graph? This is the most serious threat, and it may be enough. If every question stated its period, the canonicalizer plus a hard filter would remove the granularity trap without any ranking. The graph earns its place only on questions where the period is set by an intermediate hop, or where answering requires deliberately moving between resolutions (for example, deriving Q4 as the annual figure minus the nine-month figure). We stratify every result by whether the target period is explicit in the question. If the filter baseline matches our method on the implicit-period stratum, we will report that the problem is lexical rather than structural and drop the walk.
Why not just let an agent search with a period-filter tool? AT2QA shows that autonomous exploration beats rigid workflows on temporal QA, and this directly challenges the claim from Tian et al. that fixed centrality scaffolding helps. Our method is closer to a rigid workflow. The honest prediction is that an agent can match our accuracy but needs more calls, and that it is more prone to stopping early at the first plausible aggregate. We test both predictions with accuracy-versus-calls curves and with the granularity-confusion rate defined below. If the agent dominates at equal cost, that settles the tension in the agent's favor for this setting.
Why not just rewrite the query? RaFe trains rewriters from reranker feedback and gains 2–3% QA accuracy in open-domain settings. InteR reaches 50.0 MAP on TREC DL'19 against 41.8 for HyDE. Rewriting can turn "Q3" into "three months ended September 30." It cannot stop a retriever from ranking the adjacent nine-month column higher, because the two columns share nearly all their vocabulary. We include a RaFe-style rewriter in front of dense retrieval as a baseline. Our prediction is that it fixes recall but leaves the confusion rate unchanged.
Why not just use a long-context model on the whole filing? DocFinQA reports that retrieval-free GPT-4 does much worse than humans on filings over 100K tokens, though no reported number was found on Lacuna for exact scores. FCMR reports that o1 reaches only 43% on hard three-hop financial queries. We still run a long-context reader on single filings as a reference point.
Proposed Experiments
Diagnostic set. The main evaluation set is built from adjacent 10-Q and 10-K filings of the same companies, starting from the filings underlying DocFinQA and SEC-10Q. We write "granularity-trap" question pairs whose answers differ between the quarterly, year-to-date, and annual columns. Half of the pairs state the period explicitly. The other half define it through an intermediate fact, such as an event date taken from the text.
Because the wrong-resolution values are known, we can measure a granularity-confusion rate: the fraction of answers that equal a value from a different resolution of the same line item. We also measure an evidence resolution-mismatch rate: the fraction of top-k evidence at a resolution different from the gold evidence.
Decisive tests. The first test is whether static centrality anchoring raises the confusion rate relative to dense retrieval. If it does not, the motivating hypothesis is false and the note's premise fails. The second test is whether resolution-conditioned ranking lowers that rate on the implicit-period stratum, compared with both the period-filter ablation and the agent at matched cost.
Transfer checks. We test transfer on existing benchmarks using their standard metrics: TAT-QA F1 (target above RELOOP's 83.5), FinQA execution accuracy, the FCMR Hard tier (where Claude 3.5 Sonnet scores 30.4%), and the temporal queries of MultiHop-RAG. On those temporal queries EventRAG reports 0.8786 answer correctness. They also test whether the method helps beyond financial periods, on news timestamps at day versus week resolution.
Cost. We report graph-construction cost against HLG's proposition extraction, and LLM calls per question for every system.
References
- A Systematic Exploration of Knowledge Graph Alignment with Large Language Models in Retrieval Augmented Generation
- Decision support from financial disclosures with deep neural networks and transfer learning
- The “two sessions”: institutional investors selloff to avoid ambiguity
- EventRAG: Enhancing LLM Generation with Event Knowledge Graphs
- Hierarchical Lexical Graph for Enhanced Multi-Hop Retrieval
- FinQA: A Dataset of Numerical Reasoning over Financial Data
- TAT-QA: A Question Answering Benchmark on a Hybrid of Tabular and Textual Content in Finance
- Interleaving Retrieval with Chain-of-Thought Reasoning for Knowledge-Intensive Multi-Step Questions
- MultiHop-RAG: Benchmarking Retrieval-Augmented Generation for Multi-Hop Queries
- Hierarchical Sequence Iteration for Heterogeneous Question Answering
- Let the Agent Search: Autonomous Exploration Beats Rigid Workflows in Temporal Question Answering
- RaFe: Ranking Feedback Improves Query Rewriting for RAG
- Synergistic Interplay between Search and Large Language Models for Information Retrieval
- DocFinQA: A Long-Context Financial Reasoning Dataset
- FCMR: Robust Evaluation of Financial Cross-Modal Multi-Hop Reasoning