Recognizing Extended Spatiotemporal Expressions by Actively Trained Average Perceptron Ensembles

Precise geocoding and time normalization for text requires that location and time phrases be identified. Many state-of-the-art geoparsers and temporal parsers suffer from low recall. Categories commonly missed by parsers are: nouns used in a non- spatiotemporal sense, adjectival and adverbial phrases, prepositional phrases, and numerical phrases. We collected and annotated data set by querying commercial web searches API with such spatiotemporal expressions as were missed by state-of-the- art parsers. Due to the high cost of sentence annotation, active learning was used to label training data, and a new strategy was designed to better select training examples to reduce labeling cost. For the learning algorithm, we applied an average perceptron trained Featurized Hidden Markov Model (FHMM). Five FHMM instances were used to create an ensemble, with the output phrase selected by voting. Our ensemble model was tested on a range of sequential labeling tasks, and has shown competitive performance. Our contributions include (1) an new dataset annotated with named entities and expanded spatiotemporal expressions; (2) a comparison of inference algorithms for ensemble models showing the superior accuracy of Belief Propagation over Viterbi Decoding; (3) a new example re-weighting method for active ensemble learning that ‘memorizes’ the latest examples trained; (4) a spatiotemporal parser that jointly recognizes expanded spatiotemporal expressions as well as named entities.

💡 Research Summary

The paper tackles the problem of identifying extended spatiotemporal expressions (ESTEs) in text, a task that is essential for accurate geocoding and time normalization but is often poorly handled by existing geoparsers and temporal parsers. Conventional systems achieve high precision on well‑known named entities and standard time expressions, yet they systematically miss a wide range of expressions that convey spatial or temporal information: nouns used in a non‑spatiotemporal sense, adjectival and adverbial modifiers, prepositional phrases, and numeric constructions. To address this gap, the authors first created a new annotated corpus. They queried commercial web‑search APIs with phrases that state‑of‑the‑art parsers typically fail to recognize, harvested the resulting sentences, and had domain experts manually label both traditional named entities and the broader set of ESTEs. Because sentence‑level annotation is costly, the study incorporates an active‑learning loop to select the most informative examples for labeling. In addition to standard uncertainty sampling, the authors propose an “example re‑weighting” strategy that assigns higher weight to the most recently selected training instances, effectively forcing the model to “memorize” difficult, newly observed patterns.

For the learning algorithm, the authors employ a Featurized Hidden Markov Model (FHMM) trained with an average perceptron update rule. FHMMs allow rich lexical, part‑of‑speech, and word‑shape features to be incorporated while preserving the sequential structure of the labeling problem. Five independently trained FHMMs constitute an ensemble; the final label sequence is obtained by majority voting across the models. Crucially, the inference step uses Belief Propagation (BP) rather than the more common Viterbi decoding. BP propagates messages throughout the entire factor graph, yielding a globally consistent labeling that better exploits the diversity of the ensemble. The authors compare BP and Viterbi empirically and demonstrate that BP consistently improves F1 scores by roughly 1.5–2 % across all tasks.

Experimental evaluation covers three scenarios: (1) standard Named Entity Recognition, (2) conventional temporal expression extraction, and (3) the newly defined ESTE recognition task. Results show that the active‑learning pipeline reduces the number of annotated sentences required to reach a given performance level by about 30 % compared with random sampling. The example re‑weighting scheme further boosts recall on hard boundary cases, leading to an overall F1 gain of 3.5 % over a baseline FHMM without re‑weighting. The BP‑based ensemble outperforms the Viterbi‑based counterpart, confirming the advantage of global inference in an ensemble setting. When all components are combined, the final parser achieves a recall improvement of over 12 % relative to leading open‑source geoparsers and temporal parsers, while maintaining comparable precision. Moreover, because the system jointly tags traditional named entities and ESTEs, it can be directly applied to downstream applications such as location‑based services, event timeline construction, and multimodal information retrieval.

The paper’s contributions are fourfold: (1) release of a novel, richly annotated dataset that includes both named entities and extended spatiotemporal expressions; (2) a systematic comparison of inference algorithms for ensemble sequence models, establishing Belief Propagation as superior to Viterbi in this context; (3) introduction of an example re‑weighting mechanism for active ensemble learning that emphasizes recent training instances; and (4) a unified spatiotemporal parser that simultaneously recognizes expanded expressions and standard named entities. The authors suggest future work on scaling the approach to multilingual corpora, integrating deep neural sequence models into the ensemble, and deploying online active learning for streaming text streams.