1 Introduction↩︎

Bridging is an anaphoric phenomenon where the referent of a newly introduced entity is inferable due to its relationship with a previously introduced entity. Consider the following sentences:

. There is a house. The door is red.¹

In example [ex:opening] above, we understand the newly introduced entity “the door” to specifically be the door of the aforementioned house, due to both the sequencing of the sentences and the semantic part-whole relationship that exists between the entities of “house” and “door”. In this bridging pair, “the door” is referred to as the bridging anaphor, and “a house” is its associative antecedent. Beyond such part-whole relations, the associative relationships that give rise to bridging manifest in a variety of different ways, including relative adjectives (a dog \(\rightarrow\) a larger/different/other dog), and prototypical associations (a library \(\rightarrow\) the books). Interpreting such implicit entity relations is necessary for various downstream NLP tasks, such as question-answering, controllable natural language generation and model output factuality verification. At present, it is unclear to what extent LLM systems track such implicit entity relations.

Bridging resolution is the task of automatically detecting bridging anaphora in natural language and resolving them back to their respective associative antecedents. While the task of bridging resolution has not received as much attention as other anaphoric phenomena like coreference resolution, it has received increased attention in recent years by being included in shared task datasets [1], [2], as well as independent efforts towards developing bridging datasets and bridging resolution systems [3]. Previous bridging resolution systems have included rule-based [4], [5], neural [6], [7], and hybrid approaches [8]. However, despite these attempts, bridging resolution has remained an extremely challenging NLP task. In the End-to-end Evaluation Setting, SoTA systems do not exceed an F1 score of 40% for anaphor recognition and do not exceed an F1 score of 30% for anaphor resolution [7], [9].

Previous work has shown the identification of bridging instances to be a very difficult task, even for human annotators, due to its highly subjective nature [10]. Thus far, there has been little exploration of LLMs’ ability to reliably recognize bridging as a phenomenon. While recent work has included efforts to provide a benchmark/baseline in limited evaluation settings [11], [12], there has been no previous attempt to leverage LLMs to perform the task of bridging resolution in the End-to-end Setting. However, previous work framing bridging anaphora resolution as a QA task [13], as well as the recent bridging resolution baseline in [12], suggest that LLM-based query systems have the potential to improve upon previous bridging resolution systems.

In this paper, we present LLMBridge, the first LLM based end-to-end pipeline for bridging resolution. We evaluate our bridging resolution pipeline on 3 English datasets for referential bridging: ISNotes [14], BASHI [15], and GUMBridge [12], and we find that our pipeline surpasses previous SoTA systems in both the End-to-end and Basic bridging resolution evaluation settings. We provide code for reproducing the LLMBridge pipeline and evaluation, as well as code for preprocessing the evaluation datasets.² We additionally provide detailed error analysis on the performance of LLMBridge, investigating the varieties of bridging that are easy/difficult for LLM based systems to identify.

2 Background on Referential Bridging↩︎

[5] introduces the distinction between referential and lexical bridging as a means of describing the differences in the bridging definitions used by English bridging corpora. ISNotes, BASHI, and GUMBridge all take an information status based definition of bridging and are, as a consequence, exclusively composed of instances of referential bridging.

Referential bridging refers to truly anaphoric instances of bridging where the bridging anaphor requires an antecedent to be interpretable, as in [ex:referential]:

. She likes the house because the windows are large.

On the other hand, lexical bridging refers to lexical semantic relations between pairs of entities, such as part-whole or set-member relations, which may or may not be anaphoric, as in [ex:lexical] where the antecedent is not strictly necessary for interpretation:

. I went to the United States last month. My first stop was Washington, DC.

Note that the meaning of “Washington, DC” in the example is recoverable without needing to refer back to “the United States”, though there is a semantic part-whole relation between the two.

In this paper, we focus on the task of bridging resolution for referential bridging.³ Table 1 shows corpus statistics for the following referential bridging corpora: ISNotes, BASHI, and GUMBridge.

3 Bridging Resolution↩︎

3.1 Task Definition↩︎

Bridging resolution is the task of automatically recognizing the bridging anaphora in a text and resolving them back to their respective associative antecedents. The task of bridging resolution can be broken up and described in the following 3 subtasks:

3.1.0.1 Anaphor Recognition

Given the text of a discourse, identify the bridging anaphora present.

3.1.0.2 Anaphor Resolution

Given a bridging anaphor in a discourse, identify the associative antecedent which makes the referent of the anaphor inferable. When this subtask is done in isolation, it can also be referred to as antecedent selection.

3.1.0.3 Subtype Classification

Given a bridging anaphor and antecedent pair in a discourse, select the sub-variety of bridging (e.g. part-whole) from a set of predefined semantic relations (a specific subtype schema must be assumed).

3.2 Evaluation Settings↩︎

The evaluation of bridging resolution is commonly carried out in the following 3 settings, each of which allows for a different amount of gold mention information to be present in the input data. In Basic Bridging Resolution,⁴ the system is given gold mention information and gold bridging anaphora, and the task is to resolve each bridging anaphor back to its respective associative antecedent (also called anaphor resolution/antecedent selection). In Full Bridging Resolution, the system is given gold mention information, and the task is to both identify bridging anaphora and resolve them back to their respective associative antecedents in a discourse. In End-to-end Bridging Resolution, the task is the same as in the Full Setting, but the system is only given raw text as input.

Previous work on bridging resolution has primarily focused on the easier evaluation settings of Basic and Full bridging resolution, and less attention has been given to the more challenging End-to-end Setting, despite it being the most realistic evaluation setting. As such, in this paper we focus on providing scores in the more realistic and challenging End-to-end Setting for the English referential bridging resources: ISNotes, BASHI, and GUMBridge. In this evaluation setting, we report P/R/F1 for Anaphor Recognition and Anaphor Resolution (the joint recognition of anaphor-antecedent pair). We also report scores in an altered version of the Basic Setting, providing the gold anaphor as input, but not providing additional gold mention information. In this evaluation setting, we report Accuracy (antecedents correctly identified / total count of gold bridging anaphora). In both evaluation settings, we additionally report subtype classification scores for LLMBridge on the GUMBridge corpus. As GUMBridge allows for multiple subtype annotations, we report both Accuracy for exact match per bridging instance and P/R/F1 on predicting individual subtype annotations.

We evaluate on the designated test split of GUMBridge, and the full datasets for ISNotes and BASHI (minus 5 documents from each corpus set aside for prompt development; see Appendix 10), as they have no test splits. While we do not provide gold mention information in either of our evaluation settings, we use predicted mention and coreference information obtained from the available Stanza coref models trained on the GUM corpus⁵ [18] (for GUMBridge) and the OntoNotes corpus⁶ [19] (for ISNotes and BASHI), which predict both coreferent and singleton mentions.

4 Previous SoTA Bridging Resolution Systems↩︎

Since the rise in prominence of neural approaches following the introduction of Transformer models [20], [21], there have been a number of attempts to address the task of bridging resolution with neural systems.

[7] give an overview of the performance of recent bridging resolution systems evaluated on ISnotes and BASHI in the more challenging End-to-end Setting, in addition to the more commonly reported Full Setting. Models adapted and evaluated in these experiments include [5], [6], and [22]. Most recently, [9] present PairSpanBERT, a pretrained model for bridging resolution based on SpanBERT [23]. Extending [7] and other previous systems by leveraging PairSpanBERT, they evaluate on BASHI and ISNotes for both the End-to-end Setting and Full Setting. The system achieves near SoTA performance for the Full Setting, with SoTA performance for the End-to-end Setting.

Taking a different approach, [13] presents a system that re-frames the task of bridging resolution as a question answering task based on context. [13]’s BARQA (QA for Bridging Anaphora Resolution) system is designed to take in a segment of text and a question about an existing bridging anaphor in that text segment as inputs, and return the corresponding associative antecedent. Using this methodology, [13] reports SoTA performance on the ISNotes and BASHI corpora for the Basic Bridging Resolution Setting.

To serve as a comparison to the performance of LLMBridge, reported scores for the SoTA systems discussed above are shown in Table 2 for the End-to-end Setting, and Table 3 for the Basic Setting.

5 LLMBridge Pipeline↩︎

LLMBridge is an LLM based pipeline for the task of bridging resolution. It combines heuristic pre/post processing with backend LLM queries to create a robust bridging resolution system for referential bridging. The pipeline handles 3 subtasks of bridging resolution: (1) anaphor recognition, (2) anaphor resolution, and (3) subtype classification. Our system provides code to run the subtasks individually (as in the Basic Bridging Resolution Setting), where gold input is given for the antecedent selection and subtype classification tasks, or as a full end-to-end pipeline, where the output of one task is used as the input for the next (as in the End-to-end Setting).

5.1 Subtask Operationalization↩︎

The following describes how we operationalize the subtasks of bridging resolution to be accomplished via LLM query. In our prompt design, we follow an information status based, referential definition of bridging, where a bridging anaphor must be a newly introduced entity, and it must be inferable due to an anaphoric, associative (non-identity) relation to a previous entity in the discourse. Complete prompt templates are included in Appendix 10. The buffer/context window sizes described in the subtask operationalizations below are all a configurable part of the preprocessing.

5.1.0.1 Anaphor Recognition

For each sentence in the input text, the backend LLM is queried and asked to identify any bridging anaphora in the sentence. The query provides a definition of bridging anaphora, instructions for the anaphor recognition task, examples, the text of the sentence being asked about with up to 5 tokens of context buffer to the left and right of the sentence, and a back context of up to 150 tokens.⁷ The LLM is instructed to return a list of bridging anaphora.

5.1.0.2 Anaphor Resolution

For each bridging anaphor predicted by the system, the backend LLM is queried and asked to identify the associative antecedent. The query provides a definition of bridging anaphora, instructions for the antecedent selection task, examples, and the text of the sentence containing the predicted anaphor (marked in double curly brackets: {{ }}) with up to 150 tokens of context to the left of the sentence. The LLM is expected to return the identified antecedent entity, or if no antecedent is identified, it is instructed to return the string “no antecedent”. In the Basic Bridging Resolution Setting, the anaphor resolution subtask is run with the sentence containing the gold anaphor and 150 tokens of back context as input.

5.1.0.3 Subtype Classification (GUMBridge only)

For each bridging anaphor-antecedent pair predicted by the system (candidate anaphora where “no antecedent” is predicted are filtered out), the backend LLM is asked to classify the bridging subtype of the pair. The query provides a definition of bridging anaphora, instructions for the subtype classification task (including details on the 11 bridging subtypes in the GUMBridge annotation schema⁸), examples, and the sentence containing the anaphor with 150 tokens of back context (with the anaphor marked in double curly brackets: {{ }} and the antecedent marked in asterisks: * *). In the Basic Bridging Resolution Setting, the LLM receives the individual sentences containing the anaphor and antecedent with 10 tokens of buffer context to the left and right. The LLM is instructed to return all of the bridging subtypes applicable to the pair.

5.2 LLM Model Selection↩︎

By design, any LLM can serve as the backend to be queried in the LLMBridge pipeline. We provide scores for one closed SoTA model, Gemini-3.1-pro-preview, and two smaller open models, meta-llama/llama-3.3-70b-instruct and Qwen2.5-7B-Instruct, in order to demonstrate the competence of different sized language models on bridging resolution. We also provide scores for Qwen2.5-7B-Instruct-FT, a version of Qwen2.5-7B-Instruct which we fine-tune using low-rank adapters (LoRA; [24]) on the GUMBridge train data (213k tokens, 4k bridging instances), which is the largest English corpus for referential bridging and includes annotations for bridging subtypes. We fine-tune the model for 3 epochs using a learning rate of 3e-4, with low-rank adapters of rank 8 and a LoRA alpha of 16. We release the Qwen2.5-7B-Instruct-FT bridging fine-tuned model with the LLMBridge pipeline code. We include this model in our experiments to investigate the gains that fine-tuning can bring for smaller LLMs, even with limited data and compute.

5.3 Pre/Post Processing↩︎

To boost the performance of our pipeline, we include a number of heuristics which are based on additional linguistic annotations, including coreference/mention, part-of-speech, and dependency syntax (UD; [25]). If the additional linguistic annotations are not available, predicted annotations can be used (we provide code to enable this within our pipeline, leveraging Stanza, [26]), or the pipeline can run without the pre/post processing. We report scores using predicted annotations from Stanza during the pre/post processing for the End-to-end and Basic evaluation settings.

5.3.0.1 Adjusting Responses to Attested Mention Spans

We adjust the span of the LLM’s predicted response to match an actual, attested (or Stanza predicted) mention span for the data in order to improve precision. We adjust the predicted anaphor returned by the LLM to be the shortest existing mention span which contains the predicted anaphor. If there is no such existing mention, we check whether there are any existing mention spans which are contained within the predicted anaphor, and, if so, adjust the anaphor span to be the longest such mention. If there is still no such mention, the predicted anaphor is rejected. We also do this adjustment of the predicted span for the antecedent selection.

5.3.0.2 Suggesting Candidate Bridging Anaphora

In order to improve recall, we also use heuristics to identify possible candidates for bridging anaphora and make additional queries to the LLM backend to get specific judgments on them. The entities we highlight include entities containing comparative adjectives (e.g., “a smaller dog”), entities containing any keyword from a list of relative or temporal markers (e.g., “another”, “others”, “different”, “following”, “yesterday”, etc.), and any two-token entity where the first token is a definite determiner (e.g, “the door”).

5.3.0.3 Filtering Coreference from Predicted Bridging Anaphora

In the anaphoric, information status based definition of referential bridging, a bridging anaphor must be a newly introduced entity in the discourse. As such, any entity with a previous mention is structurally prohibited from being a bridging anaphor. Therefore, in order to improve precision on referential bridging, we filter out any predicted bridging anaphora that are subsequent mentions in a coreference chain.

6 Results↩︎

We evaluate LLMBridge on the aforementioned English referential bridging datasets: ISNotes, BASHI, and GUMBridge. We provide single run scores on the datasets for the End-to-end Bridging Resolution Setting in Table 2, and the altered Basic Bridging Resolution Setting (gold bridging anaphor, but no gold mention information) in Table 3. When available, we also include previous SoTA bridging resolution scores reported on the datasets for comparison. In Table 4, we provide bridging subtype classification scores on GUMBridge in both evaluation settings. The highest scores for the F1 and Accuracy metrics are shown in bold.

Overall, we see that leveraging Gemini-3.1-pro-preview as the LLM backend of the pipeline yields the highest scores across all datasets and evaluation settings for LLMBridge. This is expected as Gemini-3.1-pro-preview is by far the largest model tested here, though it is notable that the scores are still not high. We observe an overall increase in scores moving from the results of Qwen2.5-7B-Instruct to that of Llama-3.3-70B-Instruct to that of Gemini-3.1-pro-preview. This demonstrates the expected trend that larger, more powerful chain-of-thought (CoT) reasoning models give stronger performance on high level discourse tasks like bridging resolution.

When comparing the different evaluation settings, we expect the End-to-end Bridging Resolution Setting to be more challenging than the Basic Bridging Resolution Setting. We see this reflected in the performance of LLMBridge across the different models and datasets, with scores increasing from the End-to-end to Basic Setting. Taking the Gemini-3.1-pro-preview GUMBridge scores as an illustrative example, we can see that the accuracy score for antecedent selection in the Basic Setting is 56.6, while the F1 scores for anaphor recognition and resolution are 50.1 and 31.3 in the End-to-end Setting. The scores for other datasets and models follow the same pattern.

Comparing the best performance of LLMBridge (Gemini-3.1-pro-preview backend) to the reported scores from previous systems, we see that we achieve SoTA results on all datasets in both the End-to-end and Basic evaluation settings. In Figure 1, we provide a comparison of the LLMBridge scores (Gemini backend) to previous SoTA results on the referential bridging datasets. For BASHI, we achieve SoTA results in the Basic Setting with an antecedent selection accuracy of 56.7 (+18.0 previous SoTA) and in the End-to-end Setting with an anaphor recognition F1 of 34.5 (+2.2) and an anaphor resolution F1 of 21.1 (+2.8). For ISNotes, we achieve SoTA results in the Basic Setting with an antecedent selection accuracy of 58.1 (+8.0), and in the End-to-end Setting with an anaphor recognition F1 of 45.6 (+5.7) and an anaphor resolution F1 of 29.8 (+3.6). For GUMBridge we achieve SoTA on antecedent selection accuracy in the Basic Setting at 56.6 (+7.0). We also provide the first End-to-end bridging resolution scores on the GUMBridge dataset. The anaphor recognition and anaphor resolution F1-Scores we report for GUMBridge, 50.1 and 31.3 respectively, are the highest F1-Scores ever reported on any English bridging dataset in the End-to-end Setting.

When we compare between the LLMBridge results for Qwen2.5-7B-Instruct and Qwen2.5-7B-Instruct-FT, we see that there is a sizable increase in performance across evaluation settings and datasets. In the End-to-end Setting we see an avg. \(\Delta\) F1 across datasets of +8.0 for anaphor recognition and +7.8 for anaphor resolution. In the Basic Setting there is an avg. \(\Delta\) accuracy of +17.0 for anaphor resolution. We also see that this jump in performance is large enough to be competitive across datasets with the performance of the larger Llama-3.3-70B-Instruct in the End-to-end Setting. With regard to subtype classification (GUMBridge only), we also see an increase in performance, with a \(\Delta\) F1 of +11.4 in the End-to-end Setting and a \(\Delta\) F1 of +24.6 in the Basic Setting. This increase in performance across subtasks illustrates the utility of fine-tuning an LLM for bridging resolution task, even when the training data and compute resources are limited. If resource availability prevents simply scaling up model size, fine-tuning a smaller model can still result in meaningful gains.

Finally, looking at the GUMBridge subtype classification scores in Table 4, we see that the performance is considerably higher in the easier Basic Setting than in the harder End-to-end Setting where classification is a downstream task. As GUMBridge allows for multiple subtype annotations on an instance of bridging, we report P/R/F1 on detection of individual subtype labels. In the Basic Setting, we achieve an F1 score of 78.5. For exact match on subtype labels, we achieve an accuracy score of 76.3 in the Basic Setting. These are the first subtype classification scores for GUMBridge reporting both P/R/F1 and accuracy, and, overall, such scores indicate moderate success for the task of bridging subtype classification.

7 Error Analysis↩︎

In this section, we conduct error analysis on the predictions from the best performing configuration of the LLMBridge pipeline (Gemini-3.1-pro-preview backend). We conduct this error analysis in order to determine what sub-varieties of referential bridging remain difficult for an LLM based bridging resolution system like LLMBridge. We explore the influence of anaphor-antecedent pair distance and bridging subtype on the ability of the LLMBridge pipeline to identify instances of bridging anaphora. We additionally examine LLMBridge’s performance on subtype categorization, looking at which subtypes are easier/harder for the system to classify. As GUMBridge is the only dataset with subtype labels, the subtype related investigations exclusively analyze the predictions on GUMBridge test. Details on the GUMBridge subtype categorization schema are given in Appendix 9.

In Figure 2, we show the distribution of anaphor-antecedent distances for bridging pairs in ISNotes, BASHI, and GUMBridge grouped by whether LLMBridge correctly identified the bridging anaphor (True Positive; TP), failed to identify the bridging anaphor (False Negative; FN), or mistakenly identified a non-bridging entity as a bridging anaphor (False Positive; FP). Because the context window given to LLMBridge for anaphor detection and antecedent selection was only 150 tokens (> 90% of bridging instances in the data are < 150 tokens apart), we first filter out long distance instances of bridging more than 150 tokens apart. Looking at the density curves in Figure 2, we see that a larger part of the density curve for False Negatives covers higher token distances when compared with the curves for True Positives and False Positives. Kolmogorov-Smirnov tests comparing the FN distribution with the TP and FP distributions confirm that LLMBridge is more likely to fail to identify instances of bridging when the anaphor and antecedent are further apart (Holm adjusted p-values both < 0.01). In other words, instances of long distance bridging are more difficult for the system to identify. It is notable that this trend is apparent even when constrained to examining instances of bridging that occur within 150 tokens. We also note the FP distribution is not significantly different from the TP distribution, indicating that although FP predictions are wrong, they are being made in a reasonable distance range.

In Figure 3, we show an association plot of the residuals from a \(\chi^2\) test for gold bridging instances that were correctly identified (TP) or missed (FN) in LLMBridge’s performance on the task of anaphor recognition and the gold subtype labels of those bridging instances. Looking at the TP row of the association plot, we can see that LLMBridge is more able to identify bridging anaphora labeled with comparison-relative or entity-meronomy. The comparison-relative instances being easier to identify likely reflects the tendency for this subtype to have overt markers (e.g., comparative markers such as “other” or “another”). entity-meronomy may be easier to recognize due to the part-whole relations being more lexically inferable from the entities alone (e.g., a shady garden bed → the soil). Looking at the FP row, we can see that LLMBridge is more likely to miss bridging instances labeled with entity-associative, entity-property, or set-member. It is unsurprising that the entity-associative subtype is difficult, as associative entity relations comprise the broadest sub-variety, covering a variety of implicit relations which lack overt markers, such as relational nouns (e.g., a business → the customer), implicit arguments (e.g., a murder → the victim), and prototypical associations (e.g., a wedding → the reception). For the subtypes entity-property and set-member, the relations may also be more abstract (e.g., intangible qualities and class-instance relations) or more dependent on a wider context for understanding (i.e., it may be difficult to infer a set relation just from the anaphor itself).

In Figure 4, we provide a confusion matrix of the bridging subtype labels assigned by LLMBridge (Gemini-3.1-pro-preview backend) and the gold labels from the test set of GUMBridge. In order to have more data to analyze, we consider the predictions from Basic evaluation setting, where the gold bridging pair is provided. Looking at Figure 4, we see that while there is overlap for a variety of different subtypes, most confusions occur only 5 or less times. If we look at the none row of the gold labels axis, we can see the counts of instances where LLMBridge extraneously generated an additional label. We can see that these hallucinated labels are primarily comparison relations. And while it is somewhat common for comparison labels to co-occur with other subtype labels (e.g., a book → another one being an instance of both comparison-relative and comparison-sense), it appears that LLMBridge over-generates such cases when only a single label is actually applicable. Looking at Figure 4, we can also see that the subtypes with the greatest overlap are entity-resultative and entity-associative, with 23 instances of resultative being mistaken by the system for instances of associative. The resultative subtype is narrow in scope, focusing specifically on causal/transformational relations between entities, which are frequently seen in contexts involving product-producing processes, such as cooking/baking (e.g., some flour → the bread, [27]), but not common in most contexts. The fact that instances of this subtype are uniformly predicted as the broader associative subtype indicates that resultative likely needs to be further detailed with more examples in the subtype classification prompt.

8 Conclusion↩︎

In this paper, we present LLMBridge, the first LLM based pipeline for bridging resolution. We evaluate our bridging resolution system in the challenging End-to-end evaluation setting, along with the Basic evaluation setting. We report scores on all 3 of the referential bridging evaluation datasets for English: ISNotes, BASHI, and GUMBridge. We additionally provide the first set of scores in the End-to-End evaluation setting for the GUMBridge corpus on bridging resolution and bridging subtype classification. We achieve new SoTA results on all 3 datasets in both the End-to-end and Basic evaluation settings. We also provide error analysis of LLMBridge’s performance on the evaluation data, finding that the system struggles to identify long distance instances of bridging as well as more abstract bridging subtypes, like entity-associative. Overall, LLMBridge produces strong results on all 3 of the referential bridging corpora for English, achieving new SoTA results across the board, including the highest ever scores reported on any English dataset for anaphor recognition and anaphor resolution in the End-to-end Setting.

Limitations↩︎

This paper focuses on creating a bridging resolution pipeline for English and does not explore the cross-linguistic differences in how bridging manifests. Future work is required to determine to what extent this pipeline could directly apply or be adapted for bridging in a multilingual setting. Additionally, the design of this pipeline focuses specifically on the anaphoric, information status based definition of bridging known as referential bridging. Further adaptation of the pipeline would be necessary to increase performance on lexical bridging. Also, results from previous work are reported rather than reproduced due to incomplete or unreproducible software repositories. This means we were not able to obtain p-values for differences between our results and previous numbers.

We also acknowledge that the cost of running the LLMBridge pipeline with a SoTA commercial LLM backend (such as Gemini-3.1-pro-preview) is a significant expense, though what is an expensive SoTA model today is likely to rapidly reduce in cost, based on past experience. The cost of the evaluation conducted in this paper was approximately 500 USD, nearly all of which was for the results obtained from running LLMBridge with Gemini-3.1-pro-preview, which is a very new and expensive model. These costs limit both our ability to conduct multiple runs for evaluation and the ability for future work to reproduce our results, at least until a corresponding model becomes more affordable. While further exploration is required to find a balancing point between backend model cost and pipeline performance, results from this study indicate that LoRA fine-tuning a mid-sized LLM may be a promising direction, and the advantages of the largest available model at the time of writing are reflected in the scores we report in our results.

9 GUMBridge Bridging Subtypes↩︎

This appendix briefly details the bridging subtype varieties annotated in the GUMBridge corpus. Figure 5 shows the proportions and raw counts of the subtype annotations in the GUMBridge corpus.

9.0.0.1 comparison-relative

The anaphor is preceded by a comparative marker which implies a comparison to the antecedent (e.g., several women → other women).

9.0.0.2 comparison-sense

The type of the anaphor is omitted but inferable via comparison to the antecedent (e.g., a Chinese restaurant → the Italian one).

9.0.0.3 comparison-time

The anaphor refers to a specific time/time frame which is understandable with reference to the time/time frame expressed by the antecedent (e.g., Wednesday → yesterday).

9.0.0.4 entity-meronomy

The anaphor has a part-whole relation with the antecedent, including physical subparts, substance-portion, and regions/subsections (e.g., a house → the door).

9.0.0.5 entity-property

The anaphor is a physical or intangible property of the antecedent, such as smell, length, or style (e.g., a bouquet of roses → the scent).

9.0.0.6 entity-resultative

The anaphor is logically inferable from the antecedent. This is often the result of a transformative/product producing process, like cooking/baking (e.g., flour → the bread).

9.0.0.7 entity-associative

The anaphor is an attribute or closely associated entity of the antecedent (e.g., a library → the books).

9.0.0.8 set-member

The anaphor is an element of the antecedent set. This includes group-member and class-instance relations (e.g., several books → the mystery novel).

9.0.0.9 set-subset

The anaphor is a subset of the antecedent set (e.g., a group of students → the boys).

9.0.0.10 set-span-interval

The anaphor is a sub-span of the spatial or temporal antecedent interval (e.g, Sunday → the morning).

9.0.0.11 other

The other category is for instances which fit the information status based definition of a bridging pair but do not fall into any of the bridging subtype categories outlined above.

10 LLMBridge Pipeline Prompts↩︎

The following are examples of the prompt templates we use to query the LLM backend of LLMBridge for each bridging resolution subtask. To develop our LLM prompts for each bridging subtask, we utilized a broad definition of referential bridging that prioritizes high recall across datasets. While we explored using more specific prompts specialized for each dataset definition of bridging, the resulting lower recall did not outweigh gains in precision.⁹ As such, our finalized prompts use the more open definition of referential bridging combined with subtype explanations from the GUMBridge categorization schema and dataset specific few-shot examples for the individual subtasks.

10.0.0.1 Anaphor Recognition

10.0.0.2 Anaphor Resolution

10.0.0.3 Subtype Classification

References↩︎

---

1. Bridging anaphora are marked in bold face, and their associative antecedents are underlined.↩︎

2. Anonymized↩︎

3. The ARRAU corpus [16], [17] annotates related mentions that establish entity coherence through non-identity relations as bridging, rather than using an information status based definition. As such, ARRAU contains a mix of referential and lexical bridging, and we do not include ARRAU in our evaluation data.↩︎

4. This setting has previously been referred to just as “Bridging Resolution”. We add the designation “Basic” to differentiate it from the general task name.↩︎

5. gum-nospeakers_roberta-large-lora↩︎

6. ontonotes-singletons_roberta-large-lora↩︎

7. An analysis of the train/dev partitions has shown > 90% of associative antecedents fall within that window.↩︎

8. Details of the subtype varieties in GUMBridge are included in Appendix 9↩︎

9. During prompt development, we evaluated against a subset documents from the GUMBridge dev set (GUM_academic_librarians, GUM_conversation_grounded, GUM_court_loan, GUM_bio_emperor, GUM_voyage_athens), as well as dev documents we set aside from ISNotes (wsj_1450, wsj_1327, wsj_1232, wsj_1163, wsj_1148) and BASHI (wsj_1846, wsj_2112, wsj_1505, wsj_0242, wsj_0790).↩︎