Transit planners and dispatch algorithms usually treat demand as a known, fixed table of trips. Even "dynamic" systems estimate that table from automatic passenger counters, which only record people who actually board. When a bus is full and drives past a waiting crowd, the counter records a full bus. It does not record the people left at the curb. Measured demand is therefore censored: it is capped by the capacity that was supplied, so plans built on it keep under-supplying the trips that were already short. Riders do report these failures, but mostly in flat factual sentences such as "the 8:15 passed three stops without opening its doors" or "I've been late to work four times this month." Standard sentiment tools read these as neutral, because they contain no opinion words.
This note proposes treating implicit-sentiment extraction as a measurement instrument for that missing demand. Implicit sentiment means opinion carried by facts or idioms rather than evaluative adjectives. A model would read each complaint, tag it with a category such as "capacity," and infer a negative judgement even when none is stated. Pooled by route, stop and time window, these tags would estimate how many riders the counters missed. The corrected demand would then feed a reserve-bus stationing policy. In two sentences: counters undercount exactly where service fails, and riders' factual complaints are a signal of that undercount. We test whether recovering it with an implicit-aware sentiment model, rather than counting complaints or matching keywords, changes where reserve buses should wait.
Core Claim
The claim is narrow and falsifiable. Complaints that an implicit-aware model labels as negative and capacity- or reliability-related predict next-period capacity saturation on specific route-time cells. Those cells are where the bus reaches the counter's maximum load, the observable footprint of left-behind riders. The text should add predictive power beyond lagged passenger counts. Feeding this corrected demand into a reserve stationing policy should serve more riders than the same policy run on counts alone.
The value is expected to come mainly from proactive stationing on recurring trips, not from reacting to a single event. Complaints arrive after the failure they describe. A complaint about Tuesday's 8:15 is useful because Wednesday's 8:15 will likely fail the same way.
Relation to Prior Work
The closest system is An Online Approach to Solving Public Transit Stationing and Dispatch Problem. It formulates reserve-bus stationing and disruption dispatch as a semi-Markov decision process, meaning a sequential decision problem in which decisions happen at irregular event times. It solves this with Monte-Carlo tree search over simulated futures, using 17 million passenger-count records and 2.3 million Waze and agency incident reports from Nashville. It reports serving 2% to 7% more passengers and 40% fewer deadhead miles (miles driven empty). That paper motivates our work and supplies our decision layer. Incident text enters it only as categorical triggers, and demand enters only as counts. We keep its planner and change what the planner believes about demand.
Two other lines frame the gap.
- The real-time dynamic online feeder service with a maximum headway at mandatory stops handles streaming requests to within 6.5% of an omniscient static plan. It assumes every request arrives as a clean app booking. It names a "paradox of freedom": too much flexibility hurts future acceptance because the solver cannot see unbooked demand. Our censored-demand estimate is one concrete form of that missing foresight.
- Learn-n-Route learns implicit dispatcher preferences from stop sequences, encoded as Markov transition costs inside a routing integer program. It reads no text. When its customer base shifted abruptly at Week 53, it could only catch up after observing new trajectories. We share its goal of capturing what operators do not write into rules, but our signal is linguistic and arrives independently of the routing history.
On the language side, three papers supply the method:
- Learning Implicit Sentiment in Aspect-based Sentiment Analysis with Supervised Contrastive Pre-Training documents that roughly 30% of reviews carry sentiment through objective facts. This "objectivity trap" is precisely the complaint style we expect from riders.
- CACA is an extractive model on DeBERTaV3-large. It learns an explicit representation for absent aspects and uses cross-attention between opinion and aspect spans. It reports 64.67 F1 on the Restaurant quadruple benchmark against OTP's 62.71, and a 19% gain on the implicit-aspect, explicit-opinion subset.
- GEN-SCL-NAT is a generative T5-large model trained with supervised contrastive learning. It reports up to 2.56 F1 gain on implicit subsets.
These papers disagree on whether extractive or generative decoding handles implicit cases better. Our setting adds a stake to that disagreement. A generative model that invents a "capacity" label on a mild comment creates phantom demand, and the planner will act on it.
Measuring Your ASTE Models in The Wild constrains the design. Extraction quality drops sharply off curated domains, so we cannot assume restaurant-trained models work on transit text without in-domain annotation.
Research Program
The work has three components.
1. Transit implicit-sentiment corpus. We build a small annotated corpus of transit feedback. Sources are agency comment-line text and operator incident notes; obtaining these under a data agreement with an agency such as WeGo is a hard dependency. Each comment receives quadruples in the format of the existing benchmarks: aspect term, aspect category, opinion term, and polarity. The aspect term or opinion term may be null. Categories follow transit operations: capacity, punctuality, skipped stop, transfer, vehicle condition, and staff. We also mark which comments are purely factual, so the objectivity-trap subset can be scored separately. Around two to three thousand comments is a realistic target. Both CACA and GEN-SCL-NAT are fine-tuned on this corpus after pretraining on the Restaurant and Laptop benchmarks.
2. Censored-demand model. For each route-direction-time cell $c$ on day $t$, true demand is $d_{c,t} = b_{c,t} + \ell_{c,t}$. Here $b_{c,t}$ is the boardings the counter records, and $\ell_{c,t}$ is the unobserved left-behind count. We model $\ell$ with a latent count model in which:
- recent saturation history and lagged boardings are always inputs;
- the number of extracted negative capacity or skipped-stop quadruples attributed to cell $c$ in the preceding days enters as an additional observation.
Fitting uses the saturation events visible in the counter data as weak supervision. Cells where load hits capacity are censored observations, and cells below capacity give $\ell \approx 0$. Attributing a comment to a cell requires resolving route, stop and time mentions. We do this with a rule-based linker over the agency's schedule and report linking accuracy separately, because errors there are not sentiment errors.
3. Stationing policy. We reimplement the stationing-and-dispatch tree search on the Nashville counter and incident data, as described in the stationing paper. We then replace its demand generator with draws from the corrected demand $b + \hat{\ell}$. The planner itself is unchanged, so any difference in riders served is attributable to the demand estimate.
Baselines
The naive reference is the static plan: the agency's fixed reserve-bus locations with no reactive dispatch. It anchors how much any method moves the outcome.
The strongest existing method is the tree-search stationing policy driven by counts and categorical incidents alone, from An Online Approach to Solving Public Transit Stationing and Dispatch Problem. Its reported gains are 2% to 7% more passengers served and 40% fewer deadhead miles relative to agency practice. It is the right competitor because it already uses everything except the text. Our variant changes only the demand generator, with the same:
- simulator;
- rollout count and wall-clock budget per decision;
- historical test period;
- hyperparameter search budget.
Beating it means a statistically reliable gain in passengers served on held-out months, at no more than a few percent extra deadhead, on recurring trips that saturate.
On the language side, the reference numbers are CACA's 64.67 F1 on Restaurant and OTP's 62.71. For implicit-sentiment extraction on transit text, no reported result was found on Lacuna. We therefore report both models on our corpus, with the factual-statement subset scored separately.
Four ablations isolate the key ingredient, which is implicit-aware labeling as opposed to the mere presence of text:
- complaint volume per cell, with no sentiment at all;
- a keyword and explicit-polarity lexicon;
- the implicit-aware extractor restricted to explicit-opinion comments;
- an oracle that uses human gold quadruples, to bound how much better extraction could help.
Why not just…?
Why not just use the passenger counters, which the stationing paper already shows work well? Counters cannot see riders who never boarded. The censoring is worst exactly where reserves would help, so a counts-only model learns that saturated trips have demand equal to capacity. This is the hypothesis, not a settled fact. The counts-only planner is our main baseline, and if it matches the text-augmented planner on saturating trips, the claim fails.
Why not just count complaints per route? This is the most serious threat. Complaint volume may already carry most of the signal, and it needs no model. A cell with five complaints of any kind might be as informative as five capacity-tagged ones. However, much feedback concerns staff behaviour, cleanliness, or fares, which says nothing about missing riders. Separating capacity failures from these should matter. The volume-only ablation settles it directly. If it equals the implicit-aware variant on both saturation prediction and riders served, we will report that sentiment labeling is unnecessary and present the result as a text-volume finding.
Why not just use a keyword lexicon or an off-the-shelf polarity classifier? The factual style of transit complaints is the case such tools mishandle. The contrastive pre-training paper finds around 30% of reviews carry sentiment only through facts. CACA's 19% gain on implicit subsets shows how much surface models leave behind. Whether that gap survives in transit text is measured by scoring the lexicon on the factual subset of our corpus and propagating its labels through the same demand model.
Why not just prompt a large language model with chain-of-thought? Multi-step prompting for implicit sentiment exists, as in Application of Multiple Chain-of-Thought in Contrastive Reasoning for Implicit Sentiment Analysis. It is a reasonable labeler, and we include it as a third extractor. The concern is false positives: invented capacity tags become phantom demand. We report its precision on capacity-category labels alongside F1, because precision, not recall, drives wasted deadhead.
Proposed Experiments
Extraction. We first establish that fine-tuned extractors reach usable accuracy on transit text. We report quadruple F1 overall and on the factual subset, and capacity-category precision. Both CACA and GEN-SCL-NAT are trained on Restaurant and Laptop, then adapted. Restaurant results are reproduced as a sanity check against the 64.67 figure.
Prediction (first decisive check). We predict next-day capacity saturation per route-time cell on held-out months. The comparison is lagged counts alone against counts plus each text variant: volume, lexicon, implicit-aware extractor, and gold labels. If the implicit-aware variant does not beat volume-only here, the downstream experiment is not worth running in its sentiment-specific form.
Lead time. We stratify the prediction results by recurring versus one-off trips. Text should help mainly on recurring ones, because complaints follow the failure they describe and only recurrence makes them forward-looking.
Operations (second decisive check). We run the stationing tree search with each demand generator under matched compute on the Nashville period. We report passengers served, deadhead miles, and the share of gains on cells flagged as censored.
Extractor stakes. Generative and extractive extractors are compared on downstream deadhead. This measures whether one family's false capacity tags translate into wasted vehicle miles, which the language benchmarks alone cannot reveal.
References
- An Online Approach to Solving Public Transit Stationing and Dispatch Problem
- The real-time dynamic online feeder service with a maximum headway at mandatory stops
- Learn-n-Route: Learning implicit preferences for vehicle routing
- Learning Implicit Sentiment in Aspect-based Sentiment Analysis with Supervised Contrastive Pre-Training
- CACA: Context-Aware Cross-Attention Network for Extractive Aspect Sentiment Quad Prediction
- Generative Aspect-Based Sentiment Analysis with Contrastive Learning and Expressive Structure
- Measuring Your ASTE Models in The Wild: A Diversified Multi-domain Dataset For Aspect Sentiment Triplet Extraction
- Application of Multiple Chain-of-Thought in Contrastive Reasoning for Implicit Sentiment Analysis