Toward Complete Hospital Discharge Summarization with Abstract Meaning Representation

# Toward Complete Hospital Discharge Summarization with Abstract Meaning Representation
Source: [https://arxiv.org/html/2609.13581](https://arxiv.org/html/2609.13581)
Sitara RaoAffiliation:College of Medicine University of Illinois Chicago Chicago, USA srao11@bidmc\.harvard\.eduAffiliation:Affiliation:Department of Computer Science University of Illinois Chicago Chicago, USA plande2@uic\.eduBarbara Di EugenioAffiliation:Department of Computer Science University of Illinois Chicago Chicago, USA bdieugen@uic\.eduAaron Chaise

###### Abstract

Discharge summaries are lengthy medical documents that summarize a hospital in\-patient visit\. Automatically generating them can reduce documentation burden and return clinician time to patient care\. Whereas Large language models \(LLMs\) could be used for this task, their Achilles heel is hallucinations, which can have drastic consequences for clinical documentation\. We present an evidence\-driven alignment framework for discharge summarization at the clinical encounter level, that treats provenance as a first\-class constraint, using semantic graphs and deep learning models\. Each summary sentence is selected and organized via cross\-document semantic alignment and is accompanied by explicit evidence links to its source spans\. We show our results on two corpora: a publicly available corpus \(MIMIC\-III\) and clinical notes written by physicians at the University of Illinois Hospital\. Additionally, we make source code and trained models available\.

###### Index Terms:

summarization, EHR, graph, AMR\.

## IIntroduction

Automatic summarization is the task of using computers to summarize natural language text\[[37](https://arxiv.org/html/2609.13581#bib.bib37),[43](https://arxiv.org/html/2609.13581#bib.bib43),[8](https://arxiv.org/html/2609.13581#bib.bib8),[56](https://arxiv.org/html/2609.13581#bib.bib56)\]\. This task has developed across many areas in computational linguistics for more than six decades\[[33](https://arxiv.org/html/2609.13581#bib.bib33)\]\. Recently large language models have been influential\[[5](https://arxiv.org/html/2609.13581#bib.bib5),[48](https://arxiv.org/html/2609.13581#bib.bib48),[15](https://arxiv.org/html/2609.13581#bib.bib15)\]and shown to achieve state\-of\-the\-art performance in summarization\[[26](https://arxiv.org/html/2609.13581#bib.bib26),[53](https://arxiv.org/html/2609.13581#bib.bib55)\]\. However, given their memory constraints, these models cannot condense a large number of long documents\[[49](https://arxiv.org/html/2609.13581#bib.bib49),[32](https://arxiv.org/html/2609.13581#bib.bib32),[2](https://arxiv.org/html/2609.13581#bib.bib2)\]and lack explainability\[[55](https://arxiv.org/html/2609.13581#bib.bib54),[23](https://arxiv.org/html/2609.13581#bib.bib24)\]\. Hallucinations \(erroneous and nonfactual generated text\) present additional challenges with large language model summarizations\[[19](https://arxiv.org/html/2609.13581#bib.bib19),[16](https://arxiv.org/html/2609.13581#bib.bib16),[51](https://arxiv.org/html/2609.13581#bib.bib51)\]\.

Our goal is to automatically generate an English discharge summary using clinical notes upon discharge of a patient from a hospital\. A discharge summary is a medical document that explains a patient’s illness, reason for their hospital stay, and treatment\. Clinicians write exhaustive time\-consuming documentation during hospitalization of their patients, which can be aided by an automatic process\.[Figure1](https://arxiv.org/html/2609.13581#S1.F1)illustrates the correspondence \(provenance\) between previously written clinical notes \(note antecedents\), stored in the electronic health record system, and corresponding components of the discharge summary\.

![Refer to caption](https://arxiv.org/html/2609.13581v1/note-flow.png)Fig\. 1:Discharge Summary Generation\.Clinical notes are used as source text to automatically generate a hospital discharge summary\.Clinical summarizations must be both faithful and traceable\[[45](https://arxiv.org/html/2609.13581#bib.bib45),[55](https://arxiv.org/html/2609.13581#bib.bib54)\]\. Generating discharge summariess with these requirements highlights the difficulty of the task for admissions of extended stay patients, which have electronic health record notes that can number in the thousands111Each admission’s notes include those written during a single hospital stay\.\. Contemporary state\-of\-the\-art methods, such as fine\-tuning large language models, render the task nearly impossible with the volume of information for patients with prolonged hospitalizations\. In some cases, large language models might feasibly summarize clinical documentation on a per note basis\. However, total textual content across all free text notes of admissions easily extends past the limit of large language models’ context window\. Even the impressively large 2\-million\-token window of Gemini 2\.0\[[2](https://arxiv.org/html/2609.13581#bib.bib2)\]is not large enough for lengthy admissions222The largest admission includes 1,233 notes in the MIMIC\-III corpus\.\.

Because of these memory constraints\[[51](https://arxiv.org/html/2609.13581#bib.bib51)\], other solutions are needed for the large multi\-document automatic summarization task of generating the discharge summary\. Abstractive methods can exhibit reduced faithfulness, since they attempt at generating new text, sometimes by paraphrasing, rather than directly copying source text, as in extractive summarization\[[22](https://arxiv.org/html/2609.13581#bib.bib22),[37](https://arxiv.org/html/2609.13581#bib.bib37)\]\. They also provide no traceable means of cross\-referencing the text of summarized documents\. We chose an extractive formulation with explicit evidence linking with note section coupling so outputs remain auditable under encounter\-level multi\-document inputs\.

Previous methods have shown success at summarizing a single section\[[1](https://arxiv.org/html/2609.13581#bib.bib1)\], or one category such as physician notes\[[10](https://arxiv.org/html/2609.13581#bib.bib11)\]\. However, to the best of our knowledge, no prior work has combined episode\-level clinical encounter multi\-section discharge summarization with explicit sentence\-level provenance evidence mapping\. This motivates the extractive methods formulated in this work and provides a baseline for future abstractive summarization\.

The contributions of this work includea\)an encounter\-\-level,333An encounter is a hospitalization or visit summarized as a single event\.multi\-document discharge summarization pipelineb\)a provenance\-centered evaluation strategy aligned with sentence\-level evidence links from summary statements to source spansc\)an evidence\-linking constraint treated as a first\-class outputd\)a section\-aware organization method induced by the Source Section Model \(see[SectionIII\-E](https://arxiv.org/html/2609.13581#S3.SS5)\)e\)reusable[source code](https://github.com/uic-nlp-lab/calsum)444[https://github\.com/uic\-nlp\-lab/calsum](https://github.com/uic-nlp-lab/calsum)to reproduce our resultsf\)Medical Information Mart for Intensive Care III generated discharge summaries with physician qualitative evaluations

## IIRelated Work

Summarization is a well\-established area in natural language processing\[[31](https://arxiv.org/html/2609.13581#bib.bib31),[27](https://arxiv.org/html/2609.13581#bib.bib27),[8](https://arxiv.org/html/2609.13581#bib.bib8),[33](https://arxiv.org/html/2609.13581#bib.bib33)\]that also includes work in the clinical domain\[[55](https://arxiv.org/html/2609.13581#bib.bib54),[42](https://arxiv.org/html/2609.13581#bib.bib42)\]\. The literature is rich with examples of clinical note summarization that include both longitudinal\[[14](https://arxiv.org/html/2609.13581#bib.bib14)\], and non\-longitudinal\[[41](https://arxiv.org/html/2609.13581#bib.bib41)\]note types, two examples of mutual discipline interest\. Furthermore, the shared understanding, agreement, and acknowledgment that faithful summarization is necessary, but lacking, has been thoroughly reviewed\[[55](https://arxiv.org/html/2609.13581#bib.bib54)\]despite recent efforts to ameliorate hallucination using graph methods\[[16](https://arxiv.org/html/2609.13581#bib.bib16)\]\. Canonical extractive summarization methods have focused on ranking or selecting sentences from the source text\. Classic unsupervised approaches used graph\-centrality ranking to identify salient sentences\[[8](https://arxiv.org/html/2609.13581#bib.bib8)\]\. More recent extractive methods leverage neural architectures to model cross\-sentence relationships in a unified graph and classify sentences for extraction, as proposed by[Wang et al\. \[47\]](https://arxiv.org/html/2609.13581#bib.bib47)\.

Abstract meaning representation is a semantic representation language that describes the abstract meaning of a sentence as an acyclic graph\[[3](https://arxiv.org/html/2609.13581#bib.bib3)\]\. Recent approaches that use abstract meaning representation models as the primary data representation include work in natural language generation\[[35](https://arxiv.org/html/2609.13581#bib.bib35)\], automatic machine translation\[[4](https://arxiv.org/html/2609.13581#bib.bib4)\], and question\-answer\[[29](https://arxiv.org/html/2609.13581#bib.bib29)\]\. The well\-known work of[Liu et al\. \[31\]](https://arxiv.org/html/2609.13581#bib.bib31)used reduction methods with abstract meaning representation graphs for summarization\. In this work, the authors created a fully connected graph that was used heuristically to generate abstractive text\. This was later broadened with a more comprehensive and robust abstract meaning representation graph based realization algorithm for multi\-document summarization\[[38](https://arxiv.org/html/2609.13581#bib.bib38),[27](https://arxiv.org/html/2609.13581#bib.bib27)\]\.

Our work was inspired by the work of[Liu et al\. \[31\]](https://arxiv.org/html/2609.13581#bib.bib31)and[Liao et al\. \[27\]](https://arxiv.org/html/2609.13581#bib.bib27)on abstract meaning representation\-graph reduction\. Our method differs in that it learns and supervises cross\-document alignments using a network\-flow method\[[11](https://arxiv.org/html/2609.13581#bib.bib10)\]\. These sentence\-level provenance links from summary statements to source spans provide the traceability clinical summarization demands\. In contrast, prior graph\-reduction work built on sentence compression techniques\[[46](https://arxiv.org/html/2609.13581#bib.bib46)\], such as re\-framed commodity\-flow\[[34](https://arxiv.org/html/2609.13581#bib.bib34)\], which selects edges for inclusion within a single sentence rather than cross\-document linking\. To the best of our knowledge, no prior work has combined abstract meaning representation graphs with flow\-based \(network\-flow\) alignment for discharge summary generation; while at the same time, ensuring traceability by inferring explicit sentence\-level provenance links to source spans\.

Several recent works report abstractive and LLM\-based results on MIMIC\-derived discharge\-summary generation tasks\[[56](https://arxiv.org/html/2609.13581#bib.bib56),[39](https://arxiv.org/html/2609.13581#bib.bib39)\]\. However, our work is not directly comparable because they differ in input scope \(single\-note or fixed subsets vs\. encounter\-level multi\-document aggregation\), target definition \(full discharge summary vs\. specific sections\), preprocessing, and evaluation pipelines, and they do not enforce explicit sentence\-level provenance as an output requirement\.

## IIIMethods

Our approach consists of a pipeline of six steps, as illustrated in[Figure2](https://arxiv.org/html/2609.13581#S3.F2)\. Two corpora, the Medical Information Mart for Intensive Care III Version 1\.4\[[18](https://arxiv.org/html/2609.13581#bib.bib18)\]corpus and the UIC Health Summarization Dataset \(UICSDS\), which we collected in\-house, were used for all experiments\. The UIC Health Summarization Dataset is an IRB\-approved private dataset of 11,001 admissions and 607,872 notes, which include progress notes, radiology, ECG and additional note categories from the University of Illinois Hospital hospital\. We will now discuss the modules in[Figure2](https://arxiv.org/html/2609.13581#S3.F2)in detail\.

![Refer to caption](https://arxiv.org/html/2609.13581v1/pipeline-overview.png)Fig\. 2:Pipeline Overview\.Clinical notes are preprocessed \(top\) into summarization training data by \(a\) sectioning each note by clinical topic, \(b\) parsing each sentence of each note into an abstract meaning representation graph, and \(c\) usingCalamrto align the electronic health record note antecedents with the discharge summary\. Summarization training and testing \(bottom\) includes \(d\) matching sentences by extraction usingCalamr\[[24](https://arxiv.org/html/2609.13581#bib.bib23)\]alignments, \(e\) learning to match by note section, and \(f\) generating the discharge summary\.### III\-ASection Notes by Clinical Topic

The[spaCy](https://spacy.io/)555[https://spacy\.io](https://spacy.io/)and[scispaCy](https://allenai.github.io/scispacy/)666[https://allenai\.github\.io/scispacy/](https://allenai.github.io/scispacy/)libraries were used to tokenize, sentence chunk and tag biomedical and non\-biomedical named entities from clinical text input\. MedCAT\[[21](https://arxiv.org/html/2609.13581#bib.bib21)\]was used to link tokens to Unified Medical Language System concept unique identifiers \(CUIs\) that aid in graph aligning their text\-to\-graph concepts\. The parsed text was then segmented by clinical topic using theMedSecId\[[25](https://arxiv.org/html/2609.13581#bib.bib25)\]baseline model without modification\.MedSecId\[[25](https://arxiv.org/html/2609.13581#bib.bib25)\]is a dataset of clinical section annotations from the Medical Information Mart for Intensive Care III corpus across five note types and 50 sections on which we developed this model\.

### III\-BConstruct Admission Graph

A patient is admitted to the hospital upon entering for any administered healthcare services\. From the healthcare perspective, this admission includes what is done to the patient for the duration of the hospital stay: the admission graph is a semantic representation of it\. It is composed of two disconnected graph components: all the antecedent notes for the admission and the discharge summary\.

After notes are sectioned, they are then parsed into graphs using the clinical abstract meaning representationThymeparser\[[7](https://arxiv.org/html/2609.13581#bib.bib7)\]\. Antecedent abstract meaning representation graphs for the various notes are connected to create the source graph and the abstract meaning representation graphs of the discharge summary form the summary graph\. These two disconnected components follow the structure of the source and summary components of the bipartite graph we described inCalamr\[[24](https://arxiv.org/html/2609.13581#bib.bib23)\]\[[24](https://arxiv.org/html/2609.13581#bib.bib23)\]and in further detail in[SectionIII\-C](https://arxiv.org/html/2609.13581#S3.SS3)\. However, document nodes that represent note categories, note sections and clinical text paragraphs are used between the roots \(as a ‘‘has a’’ relation\) and their respective abstract meaning representation subgraphs as shown in[Figure3](https://arxiv.org/html/2609.13581#S3.F3)\. Afterward, we joined the abstract meaning representation parser’s selected sentence roots to their corresponding paragraph nodes and used Coreference Resolution in place of concept node merging to avoid loss of data\.

![Refer to caption](https://arxiv.org/html/2609.13581v1/adm-graph.png)Fig\. 3:Admission Graph\.An admission graph of the note antecedents \(a\), and the discharge summary \(b\) with node types by proximity too the root\.
### III\-CAlign Source and Summary

Calamr\[[24](https://arxiv.org/html/2609.13581#bib.bib23)\]\(Component ALignment for Abstract Meaning Representation\) was leveraged to find clinical notes and candidate sentences to use for summarization\. We refer the reader to our previous work\[[24](https://arxiv.org/html/2609.13581#bib.bib23)\], but we give a brief overview here\. These two graphs start as separate components that become one bipartite graph\.

Nodes are connected, as bipartite edges, if their semantic similarity’s neighborhood exceeds a threshold\. This similarity measure is calculated based on embeddings assigned to concept and attribute abstract meaning representation nodes andPropBank\[[20](https://arxiv.org/html/2609.13581#bib.bib20)\]\[[20](https://arxiv.org/html/2609.13581#bib.bib20)\]roles and role set edges\. The similarity measures are also used as the information gain across the connected graph and all subgraphs of each of the two in the max flow algorithm\[[11](https://arxiv.org/html/2609.13581#bib.bib10),[9](https://arxiv.org/html/2609.13581#bib.bib9)\]\. The assigned flow values to each bipartite edge lead to the “starvation” of low information subgraphs\. Subgraphs are effectively removed by setting low flow alignment edge capacities to zero\.

The novelty of our approach lays in clinical encounter–level supervision and organization guided byCalamr\[[24](https://arxiv.org/html/2609.13581#bib.bib23)\]alignments\. We used a network\-flow formulation\[[9](https://arxiv.org/html/2609.13581#bib.bib9)\]to induce sentence\-level alignment links across encounter documents; these links serve both as supervised training pairs and as explicit evidence\-to\-source span provenance for traceability\.

TABLE I:Graph Alignment Statistics\.Alignment and reentrancy averages by admission with the number of nodes aligned by component in the admission graph for the Medical Information Mart for Intensive Care III dataset and the UICSDS\.A total of 3,520 of the 11,957 Medical Information Mart for Intensive Care III admissions were aligned withCalamr\[[24](https://arxiv.org/html/2609.13581#bib.bib23)\]to create the Source Section Dataset \(see[SectionIII\-E](https://arxiv.org/html/2609.13581#S3.SS5)\)\.[TableI](https://arxiv.org/html/2609.13581#S3.T1)shows the average number of alignments across note antecedent and discharge summary components and the average number of reentranciess \(abstract meaning representation graph nodes with more than one parent\) per admission\[[24](https://arxiv.org/html/2609.13581#bib.bib23)\]\. The “alignable” statistics are nodes that are alignment candidates, such as concept and attribute notes\. The “aligned” statistics are those nodes with alignment edges\.

Processing very large admissions \(up to 1,233 notes in Medical Information Mart for Intensive Care III\) resulted in performance issues given the large scale of text\. Of the UIC Health Summarization Dataset sample of 607K notes, the evaluation used an aligned subset of 834 notes due to performance issues given the large scale of text \(see[SectionVII](https://arxiv.org/html/2609.13581#S7)\)\.

### III\-DMatch Sentences

Aligning the UIC Health Summarization Dataset resulted in additional challenges\. The dataset has more notes across category types compared to Medical Information Mart for Intensive Care III because the latter only includes intensive care unit notes\[[23](https://arxiv.org/html/2609.13581#bib.bib24)\]\. The consequence of this more robust note variety is that admission note counts are much higher, and therefore, take much longer to align\. There is also a higher risk of missed alignments due to a potentially higher rate of reentrancies, which lead to flow issues\[[24](https://arxiv.org/html/2609.13581#bib.bib23)\]\. Even though the Medical Information Mart for Intensive Care III alignments far outnumber the UIC Health Summarization Dataset, the UIC Health Summarization Dataset has many more reentrancies\.

The Sentence Matching Algorithm uses theCalamr\[[24](https://arxiv.org/html/2609.13581#bib.bib23)\]alignments to identify the sentences that best represent the summary\. This classification is based on the sentence\-to\-sentence information gain from the aligned graph flow network\.[Figure4](https://arxiv.org/html/2609.13581#S3.F4)shows how the source sentence,Pre\-cardiac catheterization assessment, matches with the discharge summary sentence“Coronary artery disease, status post coronary artery bypass grafting,”by creating paths through the graph from a source sentence to a summary sentence\. Each sentence connected in this way becomes a candidate\.

![Refer to caption](https://arxiv.org/html/2609.13581v1/adm-match.png)Fig\. 4:Sentence Matching\.The path \(red\) of alignment flow from the source to the summary for a single sentence\. The enlarged box shows two incoming alignment flows from the source into theheartconcept with a combined flow of 0\.905\. The green arrow represents a match candidate as a result of this alignment flow and the path to their respective sentences\. Thed24/anda56/prefixes are the abstract meaning representation concept’s instance variables\.The Sentence Matching Algorithm follows:

1. 1\.For each discharge summary sentence in the reduced graph\[[24](https://arxiv.org/html/2609.13581#bib.bib23)\], use a depth\-first search to index aligned nodes \([Figure0a](https://arxiv.org/html/2609.13581#S3.F4)\)\.
2. 2\.For each indexed node in step[1](https://arxiv.org/html/2609.13581#S3.I1.i1), traverse the alignment edge to source nodes in the note antecedent component \([Figure0b](https://arxiv.org/html/2609.13581#S3.F4)\)\.
3. 3\.Annotate aligned source nodes indexed in step[2](https://arxiv.org/html/2609.13581#S3.I1.i2)with alignment flows from discharge summary component edges \([Figure0c](https://arxiv.org/html/2609.13581#S3.F4)\)\.
4. 4\.Associate the aligned node summary annotations for each respective sentence in the source component \([Figure0d](https://arxiv.org/html/2609.13581#S3.F4)\)\.
5. 5\.Create a sentence match candidate between the source and summary sentences \([Figure0e](https://arxiv.org/html/2609.13581#S3.F4)\)\.
6. 6\.Sort the source sentences by the sum of the flow from each summary sentence\.
7. 7\.Match sentences based on the flow from each summary to source sentence\.
8. 8\.All remaining unmatched note antecedent sentences are given a stubno\-sectionlabel\.

Once the source sentences are paired with distributions of summary sentences by flow in step[6](https://arxiv.org/html/2609.13581#S3.I1.i6)each source sentence is matched with zero or more summary sentences\. In step[7](https://arxiv.org/html/2609.13581#S3.I1.i7), a source sentence is matched with the summary sentence that has the maximum flow determined by the minimum sentence flow hyperparameter\. The matched summary sentence is then eliminated as a candidate for matching with any other source sentence and the source sentences are tagged with the section of the matched summary sentence\. Upon completion, antecedent sentences are tagged with the discharge summary section to which it should be added\. For example, a sentence in a radiology antecedent note marked withBrief Hospital Coursewould be added to the discharge summary under that section during the generation process\.

### III\-ELearn Section Types

We refer to the set of notes that were successfully aligned as the Source Section Dataset\. The Sentence Matching Algorithm just described in[SectionIII\-D](https://arxiv.org/html/2609.13581#S3.SS4)was used to automatically pair sentences from note antecedents to discharge summariess of this dataset\. Each sentence pair of each admission graph was used to train a section\-assignment model \(Source Section Model\) using alignment\-induced labels\. The note counts by categories are given in note\-antecedent sentences from alignment\-derived supervision, enabling section\-aware assembly with traceable evidence links\.

TABLE II:Matched Sentence Sections\.Counts of notes per admission in the Source Section Dataset across splits \(Tr\)ain, \(Te\)st and \(Val\)idation\.This dataset was automatically generated fromCalamr\[[24](https://arxiv.org/html/2609.13581#bib.bib23)\]outputs and was used as weak supervision rather than as manually curated ground truth\. We did not perform a separate manual audit of these intermediate alignments/labels in this study; instead, we validated the pipeline at the level of its intended clinical output via human evaluation of the generated, multi\-section discharge summaries, assessing section\-wise correctness and overall readability \(see[TableIV](https://arxiv.org/html/2609.13581#S4.T4)\)\.

The selected discharge summary sections, excluding theno\-sectionlabel \(a stub used for unclassified sentences\), were based on those that were considered most necessary and beneficial for summarization by a physician authoring the note, by a clinical informatics fellow and a 4thyear medical student\. The physician\-selected discharge summary sections and their counts are given in[TableII](https://arxiv.org/html/2609.13581#S3.T2)\. Most notable is the imbalance between the section labels andno\-sectionlabel\. This high disparity leads to a terse generated discharge summary, which is explained further in[SectionV](https://arxiv.org/html/2609.13581#S5)\. However, theno\-sectionlabel preserves faithfulness by preventing forced assignment when no aligned evidence exists\.

Once the Sentence Matching Algorithm was used to assign labels to source sentences \(see[SectionIII\-D](https://arxiv.org/html/2609.13581#S3.SS4)\) a bi\-directional long\-short term memory was trained to learn the discharge summary section type of each note antecedent source sentence\. A section, such asHospital Course, was a label predicted by the model indicating that not only should the sentence be added, but to which section in the discharge summary to add it\. A label ofno\-sectionmeans the sentence is to be discarded\.

A bi\-directional long\-short term memory\[[13](https://arxiv.org/html/2609.13581#bib.bib13)\]was used for learning the sentence section classification\. The GatorTron\[[50](https://arxiv.org/html/2609.13581#bib.bib50)\]clinical embeddings, the note antecedent’s note category, and the section type were used as input features to the model\. Because of the data input size \(see[SectionIII\-D](https://arxiv.org/html/2609.13581#S3.SS4)\) the model’s static embeddings were used in place of fine\-tuning\. A fully connected linear layer was added between the bi\-directional long\-short term memory and the output layer\. The bi\-directional long\-short term memory layer had a hidden size of 500 parameters, a dropout ofp=0\.15p=0\.15, a learning rate of5×10−45\\times 10^\{\-4\}and used gradient clipping\. The model was set to train for 30 epochs and converged at 24 epochs\.

### III\-FGenerate Discharge Summary

The discharge summariess were generated using the Source Section Model \(see[SectionIII\-E](https://arxiv.org/html/2609.13581#S3.SS5)\) trained on the Source Section Dataset\. The note antecedents of the Source Section Dataset’s test set were used as input to the Source Section Model\. Sentences were added to the predicted section in the generated discharge summary or discarded if theno\-sectionlabel was predicted\.

The UIC Health Summarization Dataset was used as a development set by tuning theCalamrkkthorder neighbor set hyperparameter \(𝚲\\mathbf\{\\Lambda\}\) to include more network neighborhood semantic information\. The minimum sentence flow hyperparameter \(μ𝐬\\mathbf\{\\mu\_\{s\}\}\) was also adjusted to increase the output to 248 aligned admissions with higher quality\.

The Medical Information Mart for Intensive Care III trained summarization model yielded 133 automatically generated discharge summariess and the UIC Health Summarization Dataset model generated five\. The alignment challenges described in[SectionIII\-D](https://arxiv.org/html/2609.13581#S3.SS4), such as missing discharge summariess and GPU memory constraints, show the difficulty of hospitalization summarization\. Further discussion of these challenges are described in[SectionV](https://arxiv.org/html/2609.13581#S5)\.

Fig\. 5:Gold Discharge Summary\.The physician hand written gold University of Illinois Hospital de\-identified discharge summary\. There were no modifications to this text other than adding bold type for the headers and underlines for redacted text for readability\.DischargeSummaryby\[\*\*Doctor First Name\*\*\]\[\*\*Doctor Last Name\*\*\],MDat\[\*\*Date\*\*\]6:00AM

Author:\[\*\*Doctor First Name\*\*\]\[\*\*Doctor Last Name\*\*\],MDService:NeuroCriticalCareAuthorType:Resident

Filed:\[\*\*Date\*\*\]6:09PMDateofService:\[\*\*Date\*\*\]6:00AMStatus:Attested

Editor:\[\*\*Doctor First Name\*\*\]\[\*\*Doctor Last Name\*\*\],MD\(Resident\)Cosigner:\[\*\*Doctor First Name\*\*\]\[\*\*Doctor Last Name\*\*\],MDat\[\*\*Date\*\*\]12:45PM

Attestationsignedby\[\*\*Doctor First Name\*\*\]\[\*\*Doctor Last Name\*\*\],MDat\[\*\*Date\*\*\]12:45PMStrokeattending:Ihavereviewedtheabovedischargesummaryandagreewiththeassessment\.

PtelopedbeforeIcouldstaffthept\.

\[\*\*Hospital Name\*\*\]

DischargeSummary

Patient:\[\*\*First Name\*\*\]\[\*\*Last Name\*\*\]

AdmissionDate:\[\*\*Date\*\*\]

DischargeDate:\[\*\*Date\*\*\]

DischargeDisposition:LeftAgainstMedicalAdvice

DischargeService:Stroke

DischargeAttending:\[\*\*Doctor First Name\*\*\]\[\*\*Doctor Last Name\*\*\],MD

PrimaryDiagnosis:AcuteRmedialtemporalandinternalcapsule/thalamicstroke

Other Active Diagnoses

DiagnosisDateNotedPOA

\-Troponinlevelelevated\[\*\*Date\*\*\]Yes

Priority:High

\-Stroke\(CMS/HCC\)\[\*\*Date\*\*\]Yes

Hospital Course

HPI:\[\*\*First Name\*\*\]\[\*\*Last Name\*\*\]47y\.o\.PMHHTN,HLD,previousstroke\[\*\*Date\*\*\]\(R\-MCAwatershed\),presentingtoEDwithLsidedweakness,LLEnumbness,andblurryvisionb/l\.PatientreportssymptomsstartedacutelyonFriday\[\*\*Date\*\*\]around4pmwhiledrivingcausinghimtohavetopullover\.Hedecidedtotrytosleepitoff\.Afterwakingupthefollowingmorningwithnoimprovement,hewentto\[\*\*Hospital Name\*\*\]EDwherehewasseenbyneurology,butleftAMAashefelthewasbeingaskedthesamequestionsrepeatedlyandnothingwasgettingdone\.CTHat\[\*\*Hospital Name\*\*\]waswithoutICH,reportedlyshowedwedgeshapedhypodensityinfrontallobelikelyfromchronicinfarct\.Thepatientreportsthattheleftsidedweaknesshasimprovedsomewhattoday,buthestillendorsesb/lblurryvisionwithconstantwhitefloaters,aswellasnumbness/tinglinginhisLLE\.NIHSSis2\(LUQquadrantopia,LLEnumbness\)\.CTAHmultifocalnarrowingofb/lACAsaswellasRM1focalnarrowing\.Bw/oacuteischemiainmedialRtemporallobeinvolvingtheposteriorlimbofRinternalcapsuleaswellaspossiblythethalamus,andredemonstratingoldRfrontalACA\-MCAwatershedinfarctandoldLoccipitalcorticalinfarct\.Patienttobeadmittedtostepdownunderstrokeservice\.

PatientwithpoorlycontrolledHTNandHLD,nottakingthepast5monthshesayshehadtroublegettingprimarycareappointmentprescriptionrenewals\.SBPgetsupto200sperpatient\.LabstodaysignificantforTroponinof0\.77\-0\.62\.EKGshowingTinversionsinV5andV6\.Cardiologyconsulted\.PatientdenyingCPatthistime\.CardiorecommendingtrendingEKG/tropuntildowntrendandorderingEcho\.

Patientappearedtohaveleftbeforehewasevaluated\[\*\*Date\*\*\]AM\.

PertinentPhysicalExamAtTimeofDischarge

PhysicalExam

PATIENTNOTEXAMINEDPRIORTODISCHARGE

TestResultsPendingAtDischarge

Discharge Medications

Nomedicationshavebeenprescribed\.

IssuesRequiringFollow\-Up

\-Patientnotevaluatedpriortodischarge

OutpatientFollow\-UpAppointments

Nofutureappointments\.

Referrals

Noordersofthedefinedtypeswereplacedinthisencounter\.

Completed Consults:

ConsultsOrderedThisEncounter

Procedures

\-InpatientconsulttoGeneralNeurology

\-InpatientconsulttoCardiology

## IVEvaluation Setup

F1, precision and recall metrics were used to evaluate the source section model\. For summarization, we did not directly compare with published baselines due to differing evaluation pipelines\. Instead, our focus was evaluation for provenance for the joint purposes of faithfulness and traceability\.

Less than half of the electronic health record note text was represented in the discharge summary\[[1](https://arxiv.org/html/2609.13581#bib.bib1),[23](https://arxiv.org/html/2609.13581#bib.bib24)\]\. For additional context, we reference published discharge summary baselines of[Zhu et al\. \[56\]](https://arxiv.org/html/2609.13581#bib.bib56)\. However, we do not treat their reported scores as directly comparable due to differences in input scope, target definition, and evaluation pipelines\. Encounter\-\-level summarization is intrinsically difficult because discharge summary content often lexically differs from note antecedent text despite being semantically related\. This motivated cross\-document provenance\-centered evaluation that prioritized evidence traceability over surface overlap\.

### IV\-ALimitation of Automatic Evaluation Metrics

Automatic metrics on generated summaries, such asRouge\[[30](https://arxiv.org/html/2609.13581#bib.bib30)\],Bleu\[[40](https://arxiv.org/html/2609.13581#bib.bib40)\]andBERTScore\[[54](https://arxiv.org/html/2609.13581#bib.bib53)\], are of little help with such a large disjoint set of textual documents\. Still, we compared the electronic health record records with the discharge summary, by concatenating the note antecedents for comparison\. Automatic overlap metrics were uniformly low in this encounter\-level setting\. Specifically,Rouge1 was 27\.98%,Rouge2 was 9\.73%, andRougeL was 12\.41%;Bleu\[[40](https://arxiv.org/html/2609.13581#bib.bib40)\]was 6\.36%\. These values reflected the substantial lexical mismatch between antecedent notes and discharge summaries\. Accordingly, we treatRougeandBleu\[[40](https://arxiv.org/html/2609.13581#bib.bib40)\]as a coarse reference and emphasize human review based on traceable evidence links to source spans\.

Because encounter\-level discharge summaries often differed lexically from antecedent notes, and because published baselines and ablations varied substantially in input scope and target definition \(and typically did not enforce provenance\), we report provenance\-centered results and refer the reader to baseline papers by[Lewis et al\. \[26\]](https://arxiv.org/html/2609.13581#bib.bib26)and[Zhang et al\. \[53\]](https://arxiv.org/html/2609.13581#bib.bib55)for non\-comparable context\. For these reasons we believe human evaluation is appropriate for judging the effectiveness of generated documentation given the depth, complexity and technical jargon found in clinical notes\.

This evaluation of discharge summariess on the Source Section Dataset’s test set was evaluated by a clinical informatics fellow and a 4thyear medical student\. Each generated discharge summary was ranked using a Likert scale\[[28](https://arxiv.org/html/2609.13581#bib.bib28)\]as an integer value ranking in the range 1 to 5 with five as the highest on the following questions:a\)preference “do you prefer the generated summary”b\)readability “of the data in the generated summary, how readable is it”c\)correctness “of the data that is in the generated summary, how correct is it”d\)complete “how complete is the generated summary”e\)sections “of the data in the summary, how well is it sectioned”

### IV\-BParser Assessment

The choice of abstract meaning representation parser has the potential of greatly affecting performance\. Poor performance by the parser leads to error propagation because of semantic use of its output and because it is one of the first components of the pipeline\. Given the importance of the abstract meaning representation parser, we evaluated several parsers for correctness for the clinical domain\.

Two abstract meaning representation parsers were fine\-tuned on the Bio\-AMR corpus\[[36](https://arxiv.org/html/2609.13581#bib.bib36)\]and then judged for correctness by the clinical reviewers\. Nine graphs were scored with the T5 amrlib trained parser\[[17](https://arxiv.org/html/2609.13581#bib.bib17)\]and another nine scored with the Gsii parser\[[6](https://arxiv.org/html/2609.13581#bib.bib6)\]\. The qualitative analysis of the graphs by the clinical reviewers led to the conclusion that the generated graphs were insufficient\. An evaluation of abstract meaning representation graphs created by theThymeparser\[[7](https://arxiv.org/html/2609.13581#bib.bib7)\]demonstrated significantly improved results\. This analysis gave sufficient motivation to use theThymeparser\[[7](https://arxiv.org/html/2609.13581#bib.bib7)\]for the remainder of the experiments\.

TABLE III:Source Section Model Results\.The results as weighted, micro and macro scores of the source section model\. The results of the model trained on the Medical Information Mart for Intensive Care III corpus are given on the left and the UICSDS on the right\.TABLE IV:Human Evaluation\.The average Likert scale scores by question cateogry of 133 Medical Information Mart for Intensive Care III generated summaries are given on the left and five generated summaries UICSDS on the right\.

## VResults

Fig\. 6:Generated Discharge Summary\.A University of Illinois Hospital de\-identified automatically verbatim generated discharge summary as plain text\. There were no modifications to this text other than adding bold type for the headers and underlines for redacted text for readability\.History of present illness:

9:36AMStatus:AttestedEditor:\[\*\*Doctor First Name\*\*\]\[\*\*Doctor Last Name\*\*\],MD\(Resident\)RelatedNotes:OriginalNoteby\[\*\*Doctor First Name\*\*\]\[\*\*Doctor Last Name\*\*\],MD\(Resident\)filedat\[\*\*Date\*\*\]4:01PMCosigner:\[\*\*Doctor First Name\*\*\]\[\*\*Doctor Last Name\*\*\],MDat\[\*\*Date\*\*\]12:46PMConsultOrders1\.InpatientconsulttoGeneralNeurology\[\[\*\*Correspondence ID\*\*\]\]orderedby\[\*\*Doctor Last Name\*\*\]\[\*\*Doctor Last Name\*\*\],MDat\[\*\*Date\*\*\]0754Attestationsignedby\[\*\*Doctor First Name\*\*\]\[\*\*Doctor Last Name\*\*\],MDat\[\*\*Date\*\*\]12:46PMStrokeAttending:Ptelopedpriortobeingseen\.BlurredVisionandExtremityWeakness\[\*\*First Name\*\*\]\[\*\*Last Name\*\*\]isa47y\.o\.malewithPMHHTN,HLD,previousstroke\[\*\*Year\*\*\]\(RACA\-MCAwatershed\),presentingtoEDwithLsidedweakness,LLEnumbness,andblurryvisionb/l\.PatientreportssymptomsstartedacutelyonFriday\[\*\*Date\*\*\]around4pmwhiledrivingcausinghimtohavetopullover\.Hedecidedtotrytosleepitoff\.Afterwakingupthefollowingmorningwithnoimprovement,hewentto\[\*\*Hospital Name\*\*\]EDwherehewasseenbyneurology,butleftAMAashefelthewasbeingaskedthesamequestionsrepeatedlyandnothingwasgettingdone\.CTHat\[\*\*Hospital Name\*\*\]waswithoutICH,reportedlyshowedwedgeshapedhypodensityinfrontallobelikelyfromchronicinfarct\.Thepatientreportsthattheleftsidedweaknesshasimprovedsomewhattoday,buthestillendorsesb/lblurryvisionwithconstantwhitefloaters,aswellasnumbness/tinglinginhisLLE\.CTAHwithmultifocalnarrowingofb/lACAsaswellasRM1focalnarrowing\.MRIBw/oshowingacuteischemiainmedialRtemporallobeinvolvingtheposteriorlimbofRinternalcapsuleaswellaspossiblythethalamus,andredemonstratingoldRfrontalACA\-MCAwatershedinfarctandoldLoccipitalcorticalinfarct\.Patienttobeadmittedtostepdownunderstrokeservice\.PatientwithpoorlycontrolledHTNandHLD,nottakinganymedicationsforthepast5monthsashesayshehadtroublegettingprimarycareappointmentforprescriptionrenewals\.

Physical examination:

NIHSSis2\(LUQquadrantopia,LLEnumbness\)\.SBPgetsupto200sperpatient\.LabstodaysignificantforTroponinof0\.77\.EKGshowingTinversionsinV5andV6\.Cardiologyconsulted\.PatientdenyingCPatthistime\.Priorstroke/TIAs\(date,description\):RACA\-MCAwaterhsedstrokein\[\*\*Year\*\*\]perED,workedupat\[\*\*Hospital Name\*\*\],onDAPTVascularriskfactors:HTN,HLD,priorstrokePastMedical/Surgicalhistory:PastMedicalHistory:DiagnosisDateHypertension

The Source Section Model results are summarized in[TableIII](https://arxiv.org/html/2609.13581#S4.T3)\. The weighted F1 score of 88\.72 on the Medical Information Mart for Intensive Care III corpus shows good performance for discharge summary section classification\. However, the macro F1 was 20\.41\. The high weighted F1 largely reflected the dominantno\-sectionlabel, while macro F1 reflected difficulty on minority sections due to their sparsity\.

These results indicate strong performance on frequent sections, supporting section\-aware assembly of evidence\-linked \(provenance\) multi\-section outputs from note antecedents\. Macro F1 should therefore be interpreted primarily as a measure of minority\-section difficulty in the presence of a dominantno\-sectionlabel\. The model trained on the UIC Health Summarization Dataset shows lower results\. This might be due to the higher rate of reentrancies as shown in[TableI](https://arxiv.org/html/2609.13581#S3.T1)and discussed in[SectionIII\-C](https://arxiv.org/html/2609.13581#S3.SS3)\. The fact that Medical Information Mart for Intensive Care III is a curated dataset is the most likely reason the results are higher compared to UIC Health Summarization Dataset, which is unmodified and contains protected health information\.

[TableIV](https://arxiv.org/html/2609.13581#S4.T4)shows the quantitative evaluation of 133 generated discharge summariess trained on the Medical Information Mart for Intensive Care III corpus\. The evaluation illuminates the difficulty of the task and corroborates the low automated metrics between the gold electronic health record note antecedents with the gold discharge summary\. The generated summaries achieve a perfect correctness score \(5\), and an average readability; despite low scores on sectioning, completeness and preference, this signals the promise of graph methods with abstract meaning representation forfaithfulsummarization\.

The label imbalance in the Source Section Dataset might be attributed to the sparsity ofCalamr\[[24](https://arxiv.org/html/2609.13581#bib.bib23)\]’s alignments\. If this were the case, we could adjust the hyperparameters ofCalamr\[[24](https://arxiv.org/html/2609.13581#bib.bib23)\]to produce more sentence matches\. However, the lack of alignment could be justified by the lack of notes \(other than those from the intensive care unit department\) present in the Medical Information Mart for Intensive Care III corpus\. The misalignment could also be attributed in cases where the physician writes from personal experience with the patient that is otherwise lacking from the electronic health record notes\.

The discharge summariess produced by the model trained on the UIC Health Summarization Dataset \(see[TableIV](https://arxiv.org/html/2609.13581#S4.T4)\) show better completeness but slightly lower readability\. A higher sectioning score was given to the UIC Health Summarization Dataset despite the fact that theMedSecId\[[25](https://arxiv.org/html/2609.13581#bib.bib25)\]model was trained on Medical Information Mart for Intensive Care III\. This implies theMedSecId\[[25](https://arxiv.org/html/2609.13581#bib.bib25)\]is able to section the UIC Health Summarization Dataset notes or the Source Section Model is able to predict sections based on other factors such as better alignments\. A gold discharge summary is given in[Figure5](https://arxiv.org/html/2609.13581#S3.F5)and its generated counter\-part in[Figure6](https://arxiv.org/html/2609.13581#S5.F6)\.

## VIConclusions and Future Work

We introduced an alignment\-derived supervision strategy and section\-aware assembly for multi\-section encounter\-\-level discharge summarization\. The resulting output produced summaries with explicit sentence\-level provenance traceability to source spans, supporting faithful and auditable use in clinical documentation\. The Source Section Model showed promising results in assigning discharge summary section labels to note\-antecedent sentences using training data induced fromCalamr\[[24](https://arxiv.org/html/2609.13581#bib.bib23)\]alignments\. These findings were further supported by physician qualitative evaluations, which reported high correctness and reasonable readability\.

Shortcomings of the data used to train the models led to challenges that affected performance and brought to light certain limitations\. In the case of the Medical Information Mart for Intensive Care III data, the issue of omitted intensive care unit notes\[[23](https://arxiv.org/html/2609.13581#bib.bib24)\]led to worse summarizations\. The summaries evaluated are faithful in that only content from the source text is added to the summary\. They are traceable in how each sentence can be traced back via theCalamr\[[24](https://arxiv.org/html/2609.13581#bib.bib23)\]alignments\. Our early experiments show promise despite the difficulty of the task\. Evaluation on a broader generalizability dataset and ablations across alignment\-quality, section granularity, and long\-context tasks, remain future work\.

The utilization of aGNN\[[12](https://arxiv.org/html/2609.13581#bib.bib12),[44](https://arxiv.org/html/2609.13581#bib.bib44)\]s constructed fromCalamr\[[24](https://arxiv.org/html/2609.13581#bib.bib23)\]alignments could allow for quality abstractive summaries \(also left as future work\)\. We recommend a modified version of a graph recurrent neural network\[[52](https://arxiv.org/html/2609.13581#bib.bib52)\]that usesCalamr\[[24](https://arxiv.org/html/2609.13581#bib.bib23)\]alignments to train a graph induction model for summarization\. Leveraging the source component could mitigate the need for a large training dataset since the model would have more context for summarization\. The proposed modifications to the graph recurrent neural network would treat the source component as a Bayesian prior over the graph\-generation process using aLSTMcell for each iteration of the node and edge insertion, much like the method of[Cai and Lam \[6\]](https://arxiv.org/html/2609.13581#bib.bib6)\.

While large language models have great potential, hallucinations remain a limiting factor for automatic generation of discharge summariess\. We believe abstract meaning representation graph alignment techniques represent a promising frontier with the potential to generate accurate and interpretable summaries for clinical applications\.

## VIILimitations

Many of the admissions were too large for the constraints of the available hardware when running the experiments\. Processing very large admissions \(the largest admission has 1,233 notes in the Medical Information Mart for Intensive Care III corpus\) resulted in performance issues given the immense scale of source text\. We leave performance enhancements of the pipeline that would ameliorate these issues as future work\.

## VIIIAcknowledgments

This work was partially supported by award R01 CA225446 from the National Institutes of Health \(NIH\) and by a postdoctoral award by the Center for Health Equity using Machine Learning and Artificial Intelligence \(CHEMA\) at the University of Illinois Chicago\. We thank Jon Cai and Professor Martha Palmer at the University of Colorado Boulder for generously making theThymeparser\[[7](https://arxiv.org/html/2609.13581#bib.bib7)\]available\.

## References

- \[1\]G\. Adams, E\. Alsentzer, M\. Ketenci, J\. Zucker, and N\. Elhadad\(2021\)What’s in a Summary? Laying the Groundwork for Advances in Hospital\-Course Summarization\.InProceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies,pp\. 4794–4811\.External Links:[Document](https://dx.doi.org/10.18653/v1/2021.naacl-main.382),[Link](https://aclanthology.org/2021.naacl-main.382)Cited by:[§I](https://arxiv.org/html/2609.13581#S1.p5.1),[§IV](https://arxiv.org/html/2609.13581#S4.p2.1)\.
- \[2\]R\. Anil, S\. Borgeaud, J\. Alayrac, J\. Yu, R\. Soricut, J\. Schalkwyk, A\. M\. Dai, A\. Hauth, K\. Millican, D\. Silver, M\. Johnson, I\. Antonoglou, J\. Schrittwieser, A\. Glaese, J\. Chen, E\. Pitler, T\. Lillicrap, A\. Lazaridou, O\. Firat, J\. Molloy, M\. Isard, P\. R\. Barham, T\. Hennigan, B\. Lee, F\. Viola, M\. Reynolds, Y\. Xu, R\. Doherty, E\. Collins, C\. Meyer, E\. Rutherford, E\. Moreira, K\. Ayoub, M\. Goel, J\. Krawczyk, C\. Du, E\. Chi, H\. Cheng, E\. Ni, P\. Shah, P\. Kane, B\. Chan, and M\. Faruqui\(2024\)Gemini: A Family of Highly Capable Multimodal Models\.External Links:[Document](https://dx.doi.org/10.48550/arXiv.2312.11805),2312\.11805,[Link](http://arxiv.org/abs/2312.11805)Cited by:[§I](https://arxiv.org/html/2609.13581#S1.p1.1),[§I](https://arxiv.org/html/2609.13581#S1.p3.1)\.
- \[3\]L\. Banarescu, C\. Bonial, S\. Cai, M\. Georgescu, K\. Griffitt, U\. Hermjakob, K\. Knight, P\. Koehn, M\. Palmer, and N\. Schneider\(2013\)Abstract Meaning Representation for Sembanking\.InProceedings of the 7th Linguistic Annotation Workshop and Interoperability with Discourse,Sofia, Bulgaria,pp\. 178–186\.External Links:[Link](https://aclanthology.org/W13-2322)Cited by:[§II](https://arxiv.org/html/2609.13581#S2.p2.1)\.
- \[4\]R\. Blloshmi, R\. Tripodi, and R\. Navigli\(2020\)XL\-AMR: Enabling Cross\-Lingual AMR Parsing with Transfer Learning Techniques\.InProceedings of the 2020 Conference on Empirical Methods in Natural Language Processing \(EMNLP\),Online,pp\. 2487–2500\.External Links:[Document](https://dx.doi.org/10.18653/v1/2020.emnlp-main.195),[Link](https://aclanthology.org/2020.emnlp-main.195)Cited by:[§II](https://arxiv.org/html/2609.13581#S2.p2.1)\.
- \[5\]T\. Brown, B\. Mann, N\. Ryder, M\. Subbiah, J\. D\. Kaplan, P\. Dhariwal, A\. Neelakantan, P\. Shyam, G\. Sastry, A\. Askell, S\. Agarwal, A\. Herbert\-Voss, G\. Krueger, T\. Henighan, R\. Child, A\. Ramesh, D\. Ziegler, J\. Wu, C\. Winter, C\. Hesse, M\. Chen, E\. Sigler, M\. Litwin, S\. Gray, B\. Chess, J\. Clark, C\. Berner, S\. McCandlish, A\. Radford, I\. Sutskever, and D\. Amodei\(2020\)Language Models are Few\-Shot Learners\.Advances in Neural Information Processing Systems33,pp\. 1877–1901\.External Links:[Link](https://papers.nips.cc/paper/2020/hash/1457c0d6bfcb4967418bfb8ac142f64a-Abstract.html)Cited by:[§I](https://arxiv.org/html/2609.13581#S1.p1.1)\.
- \[6\]D\. Cai and W\. Lam\(2020\)AMR Parsing via Graph\-Sequence Iterative Inference\.InProceedings of the 58th Annual Meeting of the Association for Computational Linguistics,Online,pp\. 1290–1301\.External Links:[Document](https://dx.doi.org/10.18653/v1/2020.acl-main.119),[Link](https://aclanthology.org/2020.acl-main.119)Cited by:[§IV\-B](https://arxiv.org/html/2609.13581#S4.SS2.p2.1),[§VI](https://arxiv.org/html/2609.13581#S6.p3.1)\.
- \[7\]J\. Cai, K\. Wright\-Bettner, M\. Palmer, G\. Savova, and J\. Martin\(2024\)Adapting Abstract Meaning Representation Parsing to the Clinical Narrative – the SPRING THYME parser\.InProceedings of the 6th Clinical Natural Language Processing Workshop,T\. Naumann, A\. Ben Abacha, S\. Bethard, K\. Roberts, and D\. Bitterman \(Eds\.\),Mexico City, Mexico,pp\. 271–282\.External Links:[Document](https://dx.doi.org/10.18653/v1/2024.clinicalnlp-1.23),[Link](https://aclanthology.org/2024.clinicalnlp-1.23/)Cited by:[§III\-B](https://arxiv.org/html/2609.13581#S3.SS2.p2.1),[§IV\-B](https://arxiv.org/html/2609.13581#S4.SS2.p2.1),[§VIII](https://arxiv.org/html/2609.13581#S8.p1.1)\.
- \[8\]G\. Erkan and D\. R\. Radev\(2004\)LexRank: Graph\-based Lexical Centrality as Salience in Text Summarization\.Journal of Artificial Intelligence Research22,pp\. 457–479\.External Links:[Document](https://dx.doi.org/10.1613/jair.1523),ISSN 1076\-9757,[Link](https://www.jair.org/index.php/jair/article/view/10396)Cited by:[§I](https://arxiv.org/html/2609.13581#S1.p1.1),[§II](https://arxiv.org/html/2609.13581#S2.p1.1)\.
- \[9\]L\. R\. Ford and D\. R\. Fulkerson\(1962\)Flows in networks\.InFlows in Networks,Princeton Landmarks in Mathematics and Physics,pp\. 212\.External Links:ISBN 978\-0\-691\-65184\-2,[Link](https://press.princeton.edu/books/hardcover/9780691651842/flows-in-networks)Cited by:[§III\-C](https://arxiv.org/html/2609.13581#S3.SS3.p2.1),[§III\-C](https://arxiv.org/html/2609.13581#S3.SS3.p3.1)\.
- \[10\]Y\. Gao, D\. Dligach, T\. Miller, D\. Xu, M\. M\. M\. Churpek, and M\. Afshar\(2022\)Summarizing Patients’ Problems from Hospital Progress Notes Using Pre\-trained Sequence\-to\-Sequence Models\.InProceedings of the 29th International Conference on Computational Linguistics,Gyeongju, Republic of Korea,pp\. 2979–2991\.External Links:[Link](https://aclanthology.org/2022.coling-1.264)Cited by:[§I](https://arxiv.org/html/2609.13581#S1.p5.1)\.
- \[11\]Y\. Gao, Y\. P\. Liu, and R\. Peng\(2022\)Fully Dynamic Electrical Flows: Sparse Maxflow Faster Than Goldberg\-Rao\.In2021 IEEE 62nd Annual Symposium on Foundations of Computer Science \(FOCS\),pp\. 516–527\.External Links:[Document](https://dx.doi.org/10.1109/FOCS52979.2021.00058),ISSN 2575\-8454Cited by:[§II](https://arxiv.org/html/2609.13581#S2.p3.1),[§III\-C](https://arxiv.org/html/2609.13581#S3.SS3.p2.1)\.
- \[12\]J\. Gilmer, S\. S\. Schoenholz, P\. F\. Riley, O\. Vinyals, and G\. E\. Dahl\(2017\)Neural Message Passing for Quantum Chemistry\.InProceedings of the 34th International Conference on Machine Learning,pp\. 1263–1272\.External Links:ISSN 2640\-3498,[Link](https://proceedings.mlr.press/v70/gilmer17a.html)Cited by:[§VI](https://arxiv.org/html/2609.13581#S6.p3.1)\.
- \[13\]A\. Graves and J\. Schmidhuber\(2005\)Framewise phoneme classification with bidirectional LSTM networks\.InProceedings\. 2005 IEEE International Joint Conference on Neural Networks,Vol\.4,Montreal, Canada,pp\. 2047–2052\.External Links:[Document](https://dx.doi.org/10.1109/IJCNN.2005.1556215),ISSN 2161\-4407Cited by:[§III\-E](https://arxiv.org/html/2609.13581#S3.SS5.p5.1)\.
- \[14\]J\. S\. Hirsch, J\. S\. Tanenbaum, S\. Lipsky Gorman, C\. Liu, E\. Schmitz, D\. Hashorva, A\. Ervits, D\. Vawdrey, M\. Sturm, and N\. Elhadad\(2015\)HARVEST, a longitudinal patient record summarizer\.Journal of the American Medical Informatics Association22\(2\),pp\. 263–274\.External Links:[Document](https://dx.doi.org/10.1136/amiajnl-2014-002945),ISSN 1527\-974X, 1067\-5027,[Link](https://academic.oup.com/jamia/article/22/2/263/694965)Cited by:[§II](https://arxiv.org/html/2609.13581#S2.p1.1)\.
- \[15\]E\. J\. Hu, Y\. Shen, P\. Wallis, Z\. Allen\-Zhu, Y\. Li, S\. Wang, L\. Wang, and W\. Chen\(2021\)LoRA: Low\-Rank Adaptation of Large Language Models\.InInternational Conference on Learning Representations,External Links:[Link](https://openreview.net/forum?id=nZeVKeeFYf9)Cited by:[§I](https://arxiv.org/html/2609.13581#S1.p1.1)\.
- \[16\]D\. Iarosh, A\. Panchenko, and M\. Salnikov\(2025\)On Reducing Factual Hallucinations in Graph\-to\-Text Generation Using Large Language Models\.InProceedings of the Workshop on Generative AI and Knowledge Graphs \(GenAIK\),G\. A\. Gesese, H\. Sack, H\. Paulheim, A\. Merono\-Penuela, and L\. Chen \(Eds\.\),Abu Dhabi, UAE,pp\. 43–53\.External Links:[Link](https://aclanthology.org/2025.genaik-1.5/)Cited by:[§I](https://arxiv.org/html/2609.13581#S1.p1.1),[§II](https://arxiv.org/html/2609.13581#S2.p1.1)\.
- \[17\]B\. Jascob\(2022\)Amrlib\.External Links:[Link](https://github.com/bjascob/amrlib)Cited by:[§IV\-B](https://arxiv.org/html/2609.13581#S4.SS2.p2.1)\.
- \[18\]A\. E\. W\. Johnson, T\. J\. Pollard, L\. Shen, L\. H\. Lehman, M\. Feng, M\. Ghassemi, B\. Moody, P\. Szolovits, L\. A\. Celi, and R\. G\. Mark\(2016\)MIMIC\-III, a freely accessible critical care database\.Sci Data3\(1\),pp\. 1–9\.External Links:[Document](https://dx.doi.org/10.1038/sdata.2016.35),ISSN 2052\-4463,[Link](https://www.nature.com/articles/sdata201635/)Cited by:[§III](https://arxiv.org/html/2609.13581#S3.p1.1)\.
- \[19\]A\. T\. Kalai, O\. Nachum, S\. S\. Vempala, and E\. Zhang\(2025\)Why Language Models Hallucinate\.arXiv\.External Links:[Document](https://dx.doi.org/10.48550/arXiv.2509.04664),2509\.04664,[Link](http://arxiv.org/abs/2509.04664)Cited by:[§I](https://arxiv.org/html/2609.13581#S1.p1.1)\.
- \[20\]P\. Kingsbury and M\. Palmer\(2002\)From TreeBank to PropBank\.InProceedings of the Third International Conference on Language Resources and Evaluation \(LREC’02\),Las Palmas, Canary Islands \- Spain\.External Links:[Link](http://www.lrec-conf.org/proceedings/lrec2002/pdf/283.pdf)Cited by:[§III\-C](https://arxiv.org/html/2609.13581#S3.SS3.p2.1)\.
- \[21\]Z\. Kraljevic, T\. Searle, A\. Shek, L\. Roguski, K\. Noor, D\. Bean, A\. Mascio, L\. Zhu, A\. A\. Folarin, A\. Roberts, R\. Bendayan, M\. P\. Richardson, R\. Stewart, A\. D\. Shah, W\. K\. Wong, Z\. Ibrahim, J\. T\. Teo, and R\. J\. B\. Dobson\(2021\)Multi\-domain clinical natural language processing with MedCAT: The Medical Concept Annotation Toolkit\.Artificial Intelligence in Medicine117,pp\. 102083\.External Links:[Document](https://dx.doi.org/10.1016/j.artmed.2021.102083),ISSN 0933\-3657,[Link](https://www.sciencedirect.com/science/article/pii/S0933365721000762)Cited by:[§III\-A](https://arxiv.org/html/2609.13581#S3.SS1.p1.1)\.
- \[22\]W\. Kryscinski, B\. McCann, C\. Xiong, and R\. Socher\(2020\)Evaluating the Factual Consistency of Abstractive Text Summarization\.InProceedings of the 2020 Conference on Empirical Methods in Natural Language Processing \(EMNLP\),B\. Webber, T\. Cohn, Y\. He, and Y\. Liu \(Eds\.\),Online,pp\. 9332–9346\.External Links:[Document](https://dx.doi.org/10.18653/v1/2020.emnlp-main.750),[Link](https://aclanthology.org/2020.emnlp-main.750/)Cited by:[§I](https://arxiv.org/html/2609.13581#S1.p4.1)\.
- \[23\]P\. Landes, A\. Chaise, K\. Patel, S\. Huang, and B\. Di Eugenio\(2023\)Hospital Discharge Summarization Data Provenance\.InThe 22nd Workshop on Biomedical Natural Language Processing and BioNLP Shared Tasks,Toronto, Canada,pp\. 439–448\.External Links:[Link](https://aclanthology.org/2023.bionlp-1.41)Cited by:[§I](https://arxiv.org/html/2609.13581#S1.p1.1),[§III\-D](https://arxiv.org/html/2609.13581#S3.SS4.p1.1),[§IV](https://arxiv.org/html/2609.13581#S4.p2.1),[§VI](https://arxiv.org/html/2609.13581#S6.p2.1)\.
- \[24\]P\. Landes and B\. Di Eugenio\(2024\)CALAMR: Component ALignment for Abstract Meaning Representation\.InProceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation,Torino, Italy,pp\. 2622–2637\.External Links:[Link](https://aclanthology.org/2024.lrec-main.236)Cited by:[Fig\. 2](https://arxiv.org/html/2609.13581#S3.F2),[item 1](https://arxiv.org/html/2609.13581#S3.I1.i1.p1.1),[§III\-B](https://arxiv.org/html/2609.13581#S3.SS2.p2.1),[§III\-C](https://arxiv.org/html/2609.13581#S3.SS3.p1.1),[§III\-C](https://arxiv.org/html/2609.13581#S3.SS3.p3.1),[§III\-C](https://arxiv.org/html/2609.13581#S3.SS3.p4.1),[§III\-D](https://arxiv.org/html/2609.13581#S3.SS4.p1.1),[§III\-D](https://arxiv.org/html/2609.13581#S3.SS4.p2.1),[§III\-E](https://arxiv.org/html/2609.13581#S3.SS5.p2.1),[§V](https://arxiv.org/html/2609.13581#S5.p4.1),[§VI](https://arxiv.org/html/2609.13581#S6.p1.1),[§VI](https://arxiv.org/html/2609.13581#S6.p2.1),[§VI](https://arxiv.org/html/2609.13581#S6.p3.1)\.
- \[25\]P\. Landes, K\. Patel, S\. S\. Huang, A\. Webb, B\. Di Eugenio, and C\. Caragea\(2022\)A New Public Corpus for Clinical Section Identification: MedSecId\.InProceedings of the 29th International Conference on Computational Linguistics,Gyeongju, Republic of Korea,pp\. 3709–3721\.External Links:[Link](https://aclanthology.org/2022.coling-1.326)Cited by:[§III\-A](https://arxiv.org/html/2609.13581#S3.SS1.p1.1),[§V](https://arxiv.org/html/2609.13581#S5.p5.1)\.
- \[26\]M\. Lewis, Y\. Liu, N\. Goyal, M\. Ghazvininejad, A\. Mohamed, O\. Levy, V\. Stoyanov, and L\. Zettlemoyer\(2020\)BART: Denoising Sequence\-to\-Sequence Pre\-training for Natural Language Generation, Translation, and Comprehension\.InProceedings of the 58th Annual Meeting of the Association for Computational Linguistics,Online,pp\. 7871–7880\.External Links:[Document](https://dx.doi.org/10.18653/v1/2020.acl-main.703),[Link](https://www.aclweb.org/anthology/2020.acl-main.703)Cited by:[§I](https://arxiv.org/html/2609.13581#S1.p1.1),[§IV\-A](https://arxiv.org/html/2609.13581#S4.SS1.p2.1)\.
- \[27\]K\. Liao, L\. Lebanoff, and F\. Liu\(2018\)Abstract Meaning Representation for Multi\-Document Summarization\.InProceedings of the 27th International Conference on Computational Linguistics,Santa Fe, New Mexico, USA,pp\. 1178–1190\.External Links:[Link](https://aclanthology.org/C18-1101)Cited by:[§II](https://arxiv.org/html/2609.13581#S2.p1.1),[§II](https://arxiv.org/html/2609.13581#S2.p2.1),[§II](https://arxiv.org/html/2609.13581#S2.p3.1)\.
- \[28\]R\. Likert\(1932\)A technique for the measurement of attitudes\.\.Archives of Psychology22 140,pp\. 55–55\.Cited by:[§IV\-A](https://arxiv.org/html/2609.13581#S4.SS1.p3.1)\.
- \[29\]J\. Lim, D\. Oh, Y\. Jang, K\. Yang, and H\. Lim\(2020\)I Know What You Asked: Graph Path Learning using AMR for Commonsense Reasoning\.InProceedings of the 28th International Conference on Computational Linguistics,Barcelona, Spain \(Online\),pp\. 2459–2471\.External Links:[Document](https://dx.doi.org/10.18653/v1/2020.coling-main.222),[Link](https://aclanthology.org/2020.coling-main.222)Cited by:[§II](https://arxiv.org/html/2609.13581#S2.p2.1)\.
- \[30\]C\. Lin\(2004\)ROUGE: A Package for Automatic Evaluation of Summaries\.InText Summarization Branches Out,Barcelona, Spain,pp\. 74–81\.External Links:[Link](https://aclanthology.org/W04-1013)Cited by:[§IV\-A](https://arxiv.org/html/2609.13581#S4.SS1.p1.1)\.
- \[31\]F\. Liu, J\. Flanigan, S\. Thomson, N\. Sadeh, and N\. A\. Smith\(2015\)Toward Abstractive Summarization Using Semantic Representations\.InProceedings of the 2015 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies,Denver, Colorado,pp\. 1077–1086\.External Links:[Document](https://dx.doi.org/10.3115/v1/N15-1114),[Link](http://aclweb.org/anthology/N15-1114)Cited by:[§II](https://arxiv.org/html/2609.13581#S2.p1.1),[§II](https://arxiv.org/html/2609.13581#S2.p2.1),[§II](https://arxiv.org/html/2609.13581#S2.p3.1)\.
- \[32\]N\. F\. Liu, K\. Lin, J\. Hewitt, A\. Paranjape, M\. Bevilacqua, F\. Petroni, and P\. Liang\(2024\)Lost in the Middle: How Language Models Use Long Contexts\.Transactions of the Association for Computational Linguistics12,pp\. 157–173\.External Links:[Document](https://dx.doi.org/10.1162/tacl%5Fa%5F00638),[Link](https://aclanthology.org/2024.tacl-1.9/)Cited by:[§I](https://arxiv.org/html/2609.13581#S1.p1.1)\.
- \[33\]H\. P\. Luhn\(1958\)The Automatic Creation of Literature Abstracts\.IBM Journal of Research and Development2\(2\),pp\. 159–165\.External Links:[Document](https://dx.doi.org/10.1147/rd.22.0159),ISSN 0018\-8646Cited by:[§I](https://arxiv.org/html/2609.13581#S1.p1.1),[§II](https://arxiv.org/html/2609.13581#S2.p1.1)\.
- \[34\]T\. L\. Magnanti and L\. A\. Wolsey\(1995\)Optimal trees\.InHandbooks in Operations Research and Management Science,Network Models, Vol\.7,pp\. 503–615\.External Links:[Document](https://dx.doi.org/10.1016/S0927-0507%2805%2980126-4),[Link](https://www.sciencedirect.com/science/article/pii/S0927050705801264)Cited by:[§II](https://arxiv.org/html/2609.13581#S2.p3.1)\.
- \[35\]E\. Manning, S\. Wein, and N\. Schneider\(2020\)A Human Evaluation of AMR\-to\-English Generation Systems\.InProceedings of the 28th International Conference on Computational Linguistics,Barcelona, Spain \(Online\),pp\. 4773–4786\.External Links:[Document](https://dx.doi.org/10.18653/v1/2020.coling-main.420),[Link](https://aclanthology.org/2020.coling-main.420)Cited by:[§II](https://arxiv.org/html/2609.13581#S2.p2.1)\.
- \[36\]J\. May and J\. Priyadarshi\(2017\)SemEval\-2017 Task 9: Abstract Meaning Representation Parsing and Generation\.InProceedings of the 11th International Workshop on Semantic Evaluation \(SemEval\-2017\),S\. Bethard, M\. Carpuat, M\. Apidianaki, S\. M\. Mohammad, D\. Cer, and D\. Jurgens \(Eds\.\),Vancouver, Canada,pp\. 536–545\.External Links:[Document](https://dx.doi.org/10.18653/v1/S17-2090),[Link](https://aclanthology.org/S17-2090/)Cited by:[§IV\-B](https://arxiv.org/html/2609.13581#S4.SS2.p2.1)\.
- \[37\]J\. Maynez, S\. Narayan, B\. Bohnet, and R\. McDonald\(2020\)On Faithfulness and Factuality in Abstractive Summarization\.InProceedings of the 58th Annual Meeting of the Association for Computational Linguistics,Online,pp\. 1906–1919\.External Links:[Document](https://dx.doi.org/10.18653/v1/2020.acl-main.173),[Link](https://www.aclweb.org/anthology/2020.acl-main.173)Cited by:[§I](https://arxiv.org/html/2609.13581#S1.p1.1),[§I](https://arxiv.org/html/2609.13581#S1.p4.1)\.
- \[38\]T\. O’Gorman, M\. Regan, K\. Griffitt, U\. Hermjakob, K\. Knight, and M\. Palmer\(2018\)AMR Beyond the Sentence: the Multi\-sentence AMR corpus\.InProceedings of the 27th International Conference on Computational Linguistics,Santa Fe, New Mexico, USA,pp\. 3693–3702\.External Links:[Link](https://aclanthology.org/C18-1313)Cited by:[§II](https://arxiv.org/html/2609.13581#S2.p2.1)\.
- \[39\]K\. Pal, S\. A\. Bahrainian, L\. Mercurio, and C\. Eickhoff\(2023\)Neural Summarization of Electronic Health Records\.arXiv\.External Links:[Document](https://dx.doi.org/10.48550/arXiv.2305.15222),2305\.15222,[Link](http://arxiv.org/abs/2305.15222)Cited by:[§II](https://arxiv.org/html/2609.13581#S2.p4.1)\.
- \[40\]K\. Papineni, S\. Roukos, T\. Ward, and W\. Zhu\(2002\)Bleu: a Method for Automatic Evaluation of Machine Translation\.InProceedings of the 40th Annual Meeting of the Association for Computational Linguistics,Philadelphia, Pennsylvania, USA,pp\. 311–318\.External Links:[Document](https://dx.doi.org/10.3115/1073083.1073135),[Link](https://aclanthology.org/P02-1040)Cited by:[§IV\-A](https://arxiv.org/html/2609.13581#S4.SS1.p1.1)\.
- \[41\]R\. Pivovarov and N\. Elhadad\(2015\)Automated methods for the summarization of electronic health records\.Journal of the American Medical Informatics Association22\(5\),pp\. 938–947\.External Links:[Document](https://dx.doi.org/10.1093/jamia/ocv032),ISSN 1067\-5027,[Link](https://doi.org/10.1093/jamia/ocv032)Cited by:[§II](https://arxiv.org/html/2609.13581#S2.p1.1)\.
- \[42\]C\. Raffel, N\. Shazeer, A\. Roberts, K\. Lee, S\. Narang, M\. Matena, Y\. Zhou, W\. Li, and P\. J\. Liu\(2020\)Exploring the limits of transfer learning with a unified text\-to\-text transformer\.J\. Mach\. Learn\. Res\.21\(1\),pp\. 140:5485–140:5551\.External Links:ISSN 1532\-4435,[Link](https://jmlr.org/papers/volume21/20-074/20-074.pdf)Cited by:[§II](https://arxiv.org/html/2609.13581#S2.p1.1)\.
- \[43\]N\. S\. Ranjitha and J\. S\. Kallimani\(2017\)Abstractive multi\-document summarization\.In2017 International Conference on Advances in Computing, Communications and Informatics \(ICACCI\),Manipal, Karnataka, India,pp\. 1690–1694\.External Links:[Document](https://dx.doi.org/10.1109/ICACCI.2017.8126086)Cited by:[§I](https://arxiv.org/html/2609.13581#S1.p1.1)\.
- \[44\]F\. Scarselli, M\. Gori, A\. C\. Tsoi, M\. Hagenbuchner, and G\. Monfardini\(2009\)The Graph Neural Network Model\.IEEE Transactions on Neural Networks20\(1\),pp\. 61–80\.External Links:[Document](https://dx.doi.org/10.1109/TNN.2008.2005605),ISSN 1941\-0093,[Link](https://ieeexplore.ieee.org/document/4700287)Cited by:[§VI](https://arxiv.org/html/2609.13581#S6.p3.1)\.
- \[45\]H\. Shing, C\. Shivade, N\. Pourdamghani, P\. Resnik, D\. Oard, and P\. Bhatia\(2021\)Towards Clinical Encounter Summarization: Learning to Compose Discharge Summaries from Prior Notes\.arXiv:2104\.13498 \[cs\]\.External Links:2104\.13498,[Link](http://arxiv.org/abs/2104.13498)Cited by:[§I](https://arxiv.org/html/2609.13581#S1.p3.1)\.
- \[46\]K\. Thadani and K\. McKeown\(2013\)Sentence Compression with Joint Structural Inference\.InProceedings of the Seventeenth Conference on Computational Natural Language Learning,Sofia, Bulgaria,pp\. 65–74\.External Links:[Link](https://aclanthology.org/W13-3508)Cited by:[§II](https://arxiv.org/html/2609.13581#S2.p3.1)\.
- \[47\]D\. Wang, P\. Liu, Y\. Zheng, X\. Qiu, and X\. Huang\(2020\)Heterogeneous Graph Neural Networks for Extractive Document Summarization\.InProceedings of the 58th Annual Meeting of the Association for Computational Linguistics,Online,pp\. 6209–6219\.External Links:[Document](https://dx.doi.org/10.18653/v1/2020.acl-main.553),[Link](https://aclanthology.org/2020.acl-main.553)Cited by:[§II](https://arxiv.org/html/2609.13581#S2.p1.1)\.
- \[48\]X\. Wang, J\. Wei, D\. Schuurmans, Q\. V\. Le, E\. H\. Chi, S\. Narang, A\. Chowdhery, and D\. Zhou\(2022\)Self\-Consistency Improves Chain of Thought Reasoning in Language Models\.InProceedings of the Eleventh International Conference on Learning Representations,Kigali Rwanda\.External Links:[Link](https://openreview.net/pdf?id=1PL1NIMMrw)Cited by:[§I](https://arxiv.org/html/2609.13581#S1.p1.1)\.
- \[49\]Z\. Xu, S\. Zhu, J\. Wang, J\. Wang, B\. Athiwaratkun, C\. Wang, J\. Zou, and C\. Zhang\(2025\)When Does Divide and Conquer Work for Long Context LLM? A Noise Decomposition Framework\.arXiv\.External Links:[Document](https://dx.doi.org/10.48550/arXiv.2506.16411),2506\.16411,[Link](http://arxiv.org/abs/2506.16411)Cited by:[§I](https://arxiv.org/html/2609.13581#S1.p1.1)\.
- \[50\]X\. Yang, A\. Chen, N\. PourNejatian, H\. C\. Shin, K\. E\. Smith, C\. Parisien, C\. Compas, C\. Martin, M\. G\. Flores, Y\. Zhang, T\. Magoc, C\. A\. Harle, G\. Lipori, D\. A\. Mitchell, W\. R\. Hogan, E\. A\. Shenkman, J\. Bian, and Y\. Wu\(2022\)GatorTron: A Large Clinical Language Model to Unlock Patient Information from Unstructured Electronic Health Records\.arXiv\.External Links:[Document](https://dx.doi.org/10.48550/arXiv.2203.03540),2203\.03540,[Link](http://arxiv.org/abs/2203.03540)Cited by:[§III\-E](https://arxiv.org/html/2609.13581#S3.SS5.p5.1)\.
- \[51\]Y\. Yehuda, I\. Malkiel, O\. Barkan, J\. Weill, R\. Ronen, and N\. Koenigstein\(2024\)InterrogateLLM: Zero\-Resource Hallucination Detection in LLM\-Generated Answers\.InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\),L\. Ku, A\. Martins, and V\. Srikumar \(Eds\.\),Bangkok, Thailand,pp\. 9333–9347\.External Links:[Document](https://dx.doi.org/10.18653/v1/2024.acl-long.506),[Link](https://aclanthology.org/2024.acl-long.506/)Cited by:[§I](https://arxiv.org/html/2609.13581#S1.p1.1),[§I](https://arxiv.org/html/2609.13581#S1.p4.1)\.
- \[52\]J\. You, R\. Ying, X\. Ren, W\. Hamilton, and J\. Leskovec\(2018\)GraphRNN: Generating Realistic Graphs with Deep Auto\-regressive Models\.InProceedings of the 35th International Conference on Machine Learning,Stockholm, Sweden,pp\. 5708–5717\.External Links:ISSN 2640\-3498,[Link](https://proceedings.mlr.press/v80/you18a.html)Cited by:[§VI](https://arxiv.org/html/2609.13581#S6.p3.1)\.
- \[53\]J\. Zhang, Y\. Zhao, M\. Saleh, and P\. Liu\(2020\)PEGASUS: Pre\-training with Extracted Gap\-sentences for Abstractive Summarization\.InProceedings of the 37th International Conference on Machine Learning,pp\. 11328–11339\.External Links:ISSN 2640\-3498,[Link](https://proceedings.mlr.press/v119/zhang20ae.html)Cited by:[§I](https://arxiv.org/html/2609.13581#S1.p1.1),[§IV\-A](https://arxiv.org/html/2609.13581#S4.SS1.p2.1)\.
- \[54\]T\. Zhang, V\. Kishore, F\. Wu, K\. Q\. Weinberger, and Y\. Artzi\(2020\)BERTScore: Evaluating Text Generation with BERT\.InProceedings of the 8th International Conference on Learning Representations,Addis Ababa, Ethopia\.External Links:[Link](https://openreview.net/forum?id=SkeHuCVFDr)Cited by:[§IV\-A](https://arxiv.org/html/2609.13581#S4.SS1.p1.1)\.
- \[55\]Y\. Zhang, D\. Merck, E\. Tsai, C\. D\. Manning, and C\. Langlotz\(2020\)Optimizing the Factual Correctness of a Summary: A Study of Summarizing Radiology Reports\.InProceedings of the 58th Annual Meeting of the Association for Computational Linguistics,pp\. 5108–5120\.External Links:[Document](https://dx.doi.org/10.18653/v1/2020.acl-main.458),[Link](https://aclanthology.org/2020.acl-main.458)Cited by:[§I](https://arxiv.org/html/2609.13581#S1.p1.1),[§I](https://arxiv.org/html/2609.13581#S1.p3.1),[§II](https://arxiv.org/html/2609.13581#S2.p1.1)\.
- \[56\]Y\. Zhu, X\. Yang, Y\. Wu, and W\. Zhang\(2023\)Leveraging Summary Guidance on Medical Report Summarization\.IEEE Journal of Biomedical and Health Informatics27\(10\),pp\. 5066–5075\.External Links:[Document](https://dx.doi.org/10.1109/JBHI.2023.3304376),ISSN 2168\-2208,[Link](https://ieeexplore.ieee.org/abstract/document/10214623)Cited by:[§I](https://arxiv.org/html/2609.13581#S1.p1.1),[§II](https://arxiv.org/html/2609.13581#S2.p4.1),[§IV](https://arxiv.org/html/2609.13581#S4.p2.1)\.