Some search tasks ask for sources rather than documents: the accounts, feeds or blogs that cover a topic steadily over months, not the one viral post that happened to match the query. The standard test of steadiness is a persistence score, which counts how evenly an author's on-topic posts spread across time windows. That score has a quiet assumption built in. "On-topic" is judged against a fixed set of query words. When a community's vocabulary turns over, as it does very quickly on social platforms, a true authority moves to the new hashtags and slang. Under the fixed word set, that author looks like someone who lost interest. Our claim is that persistence should be measured against a topic vocabulary that is rebuilt in every time window. A new term should join that vocabulary only when, in the same window, posts use it alongside terms already accepted as on-topic. The community's own bridging posts then do the translation, and the vocabulary cannot wander freely.
The opportunity comes from two findings that have never been put together. Source-ranking work in the blog-distillation literature built persistence scores over static vocabularies. Work on evolving language found that 78.3% of the top 40k hashtags on Twitter changed between 2013 and 2019, against 37.5% for ordinary words. The short-text setting makes this mismatch worse, because a single tweet carries almost no evidence on its own. Persistence summed over an author's timeline is one of the few ways to regain statistical strength. That only works if the summing does not miss the author's newer posts. Knowledge obsolescence is the lens for evaluation: a ranked list of sources is itself knowledge that goes stale. We will measure how quickly rankings built at one time decay when judged later, and whether drift-aware persistence slows that decay.
Core Claim and Mechanism
Consider a topic query $q$, time windows $w = 1, \dots, W$, and a topic vocabulary $V_w$ for each window. $V_1$ is the query plus terms that co-occur strongly with it in window 1. A candidate term $t$ enters $V_w$ by bridging. Its bridge score is the fraction of window-$w$ posts containing $t$ that also contain at least one term of $V_{w-1}$. That fraction must exceed what random co-occurrence at $t$'s frequency would produce. There is also an anchor constraint: a term must keep a minimum co-occurrence with the original query terms, summed over all windows so far. Hashtags and burst-detected emerging strings are kept as whole tokens and never split into subword pieces.
Short posts have essentially binary term counts, so evidence is counted per post rather than per term. Let $n_a(w)$ be the number of author $a$'s posts in window $w$ that match $V_w$, and let $p_a(w) = n_a(w) / \sum_{w'} n_a(w')$. The persistence score is the normalized entropy
$$P(a, q) = \frac{-\sum_{w} p_a(w) \log p_a(w)}{\log W},$$
which equals 1 for perfectly even coverage and 0 for a single burst. This is the entropy form of the Recurring-Degree metric. The only change is that each window is matched against its own $V_w$ instead of a fixed $V$. The final ranking interpolates $P(a,q)$ with an author-level relevance score. The interpolation weight is tuned on held-out topics.
Relation to Prior Work
The source-distillation literature supplies the scaffolding. ICTNET at Blog Track TREC 2010 introduced the entropy-based Recurring-Degree score to separate bursts from sustained interest on Blogs08, using fixed query terms. Language Modeling Approaches to Blog Postand Feed Finding found that publication frequency over a long timeline signals authority better than recency. Blog feed search with a post index built a two-stage system: a cheap post-level filter, then an author-level language model. It cut post-to-blog computations by 73% and reached MAP 0.3556 on Blog06 2007 topics.
That last paper also constrains our design. Its length-based pruning keeps the longest posts as the best evidence. That is impossible when every post is capped in length, so we replace length with post-level matching against the bridged vocabulary. Retrieval and Feedback Models for Blog Distillation found that single "large document" feed representations with Wikipedia anchor-text expansion beat sampling methods. It also found that ordinary feedback expansion pulls in generic boilerplate. That warning is why our expansion is anchored. Its expansion source is a static Wikipedia snapshot. Our expansion source is the stream itself, read window by window.
Three short-text papers motivate pooling evidence at the author level. Author Model and Negative Feedback Methods on TREC 2011 Microblog Track pooled an author's timeline into a topical profile but kept term statistics frozen. Finding Topic-Related Tweets Using Conversational Thread and A Hyperlink-Extended Language Model for Microblog Retrieval add context from reply threads or linked pages, at a latency cost. Our bridge statistics come from the stream itself and need no outside fetching. Simple Attention-Based Representation Learning for Ranking Short Social Media Posts is a cheap, strong neural ranker for individual tweets. It is a candidate for the post-level filter, not a competitor at the author level.
Work on evolving vocabularies supplies the evidence that fixed vocabularies break. Dynamic Language Models for Continuously Evolving Content measured hashtag turnover and introduced swapping stale tokens for emerging ones. It showed that keeping hashtags whole raised Micro-F1 from 0.314 to 0.561, but only on post-level classification. TimeLMs: Diachronic Language Models from Twitter retrains RoBERTa every quarter and reports roughly 10% degradation after one year for a model that is not updated. It also keeps a fixed tokenizer. AURORA adapts retrieval indices to lexical drift, but evaluates on synthetic token replacement. None of these three works aggregates over an author's history to judge whether that author is a source. We do, and we evaluate on organic drift.
Finally, the obsolescence literature motivates the evaluation lens. RealTime QA attributes 34% of errors to retrieval failure. SituatedQA shows that simply refreshing the retrieval corpus still leaves a 15-point accuracy deficit on facts that changed. Both papers point to the curation of the evidence pool as the weak link. A source list that stays current is one way to curate that pool. Testing that application is left for later; here we measure it only as ranking decay.
Research Program
We will build three components. The first is the window-wise vocabulary builder: burst detection for emerging whole tokens, bridge scoring, and the anchor constraint. The second is a two-stage source ranker in the style of Weerkamp et al. Its first stage retrieves posts against $V_w$. Its second stage aggregates matched posts per author and combines relevance with $P(a,q)$.
The third component is a new author-level test collection for microblogs, since the file contains no existing benchmark for ranking microblog sources over multiple years. We will assemble a multi-year public microblog archive and choose about 50 topics whose vocabulary is known to have turned over, such as epidemic-era terms of the kind mined by COVID-19 Public Opinion and Emotion Monitoring System Based on Time Series Thermal New Word Mining. Candidate authors will be pooled from all systems under test. Assessors will judge each candidate from a sample of their timeline drawn evenly across time, so that late-period posts are not missed.
We will also measure turnover for each source: the overlap between the topical terms in the first and second halves of the source's timeline. This lets every result be broken down by how much a source's language drifted.
Baselines
The naive reference point is the Blogger model with no persistence term: each author's posts concatenated into one document and ranked with query likelihood. This shows what persistence adds at all.
The strongest existing method at the source level is the two-stage Title+Full system of Blog feed search with a post index. It reports MAP 0.3556 on the Blog06 2007 topics, against roughly 0.31 for single-stage baselines. We will reimplement it, reproduce that number as a sanity check, and then run it unchanged on Blogs08 and the new microblog collection. Length pruning will be disabled where posts are capped.
The strongest persistence method is static Recurring-Degree from ICTNET at Blog Track TREC 2010. No reported result was found in the file for its MAP or P@10 on Blogs08, so we will report our reimplementation.
We will also run the large-document model with Wikipedia expansion from Retrieval and Feedback Models for Blog Distillation. Its Wikipedia snapshot will be matched to the start of each collection.
Two neural competitors will compute per-window post-to-query similarity in place of term matching. The first uses a static dense encoder that splits rare words into subword pieces. The second uses quarterly-updated TimeLMs-style encoders. Both feed the same entropy and aggregation. For no-persistence comparisons, the neural ranker of Simple Attention-Based Representation Learning for Ranking Short Social Media Posts serves as the post-level filter. That paper reports AP 0.4522 on TREC Microblog 2014, but that is post-level ranking, and no reported result was found on Lacuna for author-level microblog distillation.
All systems will share the same time windows, the same candidate pools, and the same tuning budget per system. That budget is a grid search over interpolation weights and expansion depth on the same held-out topics.
Three ablations isolate the key ingredient:
- Per-window expansion using ordinary feedback, with no bridge requirement and no anchor. This tests whether bridging matters or whether simply re-expanding each window is enough.
- Bridging with hashtags split into subword pieces. This tests the whole-token choice.
- A frozen $V_1$ applied to every window. This is static persistence with our evidence counting.
Beating the field means a statistically significant MAP gain (paired randomization test, $p<0.05$) over both the two-stage system and static Recurring-Degree on the microblog collection. That gain must be concentrated in the high-turnover third of relevant sources.
Why not just…?
Why not just use a dense encoder, which already handles paraphrase? Static encoders split new hashtags into fragments. Dynamic Language Models for Continuously Evolving Content shows that splitting costs heavily on hashtag tasks (0.314 versus 0.561 Micro-F1). TimeLMs shows that frozen models degrade over a year. Still, gradual slang shifts may be within reach of an encoder. The breakdown by turnover will show whether the static encoder loses specifically on high-turnover sources.
Why not just re-expand the query independently in each window with pseudo-relevance feedback? Pseudo-relevance feedback adds terms from the top-ranked results of an initial search. This is the most serious threat, because it is simple and adapts to each window. Elsas et al. found that unanchored feedback on social text drifts toward generic terms. With few matched short posts per window, we expect that drift to compound. Ablation 1 is exactly this alternative. If it matches the bridged version, we will report that bridging is unnecessary and that per-window feedback is the contribution.
Why not just use continually retrained language models instead of lexical bridging? Quarterly-updated encoders may capture drift as well as bridging does, or better. That would turn our method into a cheap approximation. We will report effectiveness against compute, using the continual-training step counts in Hombaiah et al. as a guide to retraining cost. If updated encoders win clearly at affordable cost, the claim narrows to the setting where retraining is unavailable.
Why not just treat each source as one big document and skip persistence? One concatenated document contains every term the author ever used, so it does capture drifted terms. But it cannot tell a sustained author from someone who posted one burst; the virality problem is what motivated Recurring-Degree. The naive baseline and the large-document baseline measure this directly. A burst-heavy subset of topics, selected by the volume spikes that ICTNET's entropy is designed to flag, will show whether persistence is needed at all.
Decisive Checks
The first check is the turnover breakdown. If drift-aware persistence beats frozen-$V_1$ persistence only on low-turnover sources, the mechanism is wrong.
The second check is decay over time. Rankings are built from data up to time $T$ and judged against assessments made at $T + 6$, $T + 12$ and $T + 24$ months. We will compare the slope of MAP decay across methods. The prediction is that bridged persistence decays more slowly than the static and subword-splitting variants, which are expected to resemble the roughly 10% one-year loss reported for frozen models.
The third check is a manual audit of 200 terms accepted into $V_w$, measuring precision of on-topic admissions with and without the anchor. This tests directly whether bridging admits real successor terms or noise.
The fourth check uses Blogs08 with a title-only index, following Weerkamp et al.'s lean index. This gives a short-text condition that has existing relevance judgments. Gains should be smaller there, since blog-era drift is milder. Gains no larger than on the microblog collection would be consistent with the turnover account; a larger gain on Blogs08 would contradict it.
References
- ICTNET at Blog Track TREC 2010
- Language Modeling Approaches to Blog Postand Feed Finding
- Blog feed search with a post index
- Retrieval and Feedback Models for Blog Distillation
- Author Model and Negative Feedback Methods on TREC 2011 Microblog Track
- Finding Topic-Related Tweets Using Conversational Thread
- A Hyperlink-Extended Language Model for Microblog Retrieval
- Simple Attention-Based Representation Learning for Ranking Short Social Media Posts
- Dynamic Language Models for Continuously Evolving Content
- TimeLMs: Diachronic Language Models from Twitter
- AURORA: Neuro-Symbolic Continual Indexing for Evolving RAG Systems
- RealTime QA: What's the Answer Right Now?
- SituatedQA: Incorporating Extra-Linguistic Contexts into QA
- COVID-19 Public Opinion and Emotion Monitoring System Based on Time Series Thermal New Word Mining