ParseJargon: Real-Time Personalized Jargon Support
- ParseJargon is an interactive system that provides real-time, personalized jargon identification and explanation in online meetings.
- It employs a two-stage inference pipeline combining live transcript analysis with user-specific filtering to enhance meeting comprehension.
- User studies show that personalized jargon support significantly boosts understanding and engagement while reducing cognitive overload.
Searching arXiv for papers related to ParseJargon and personalized jargon support. ParseJargon is an interactive LLM-powered system for real-time personalized jargon identification and explanation in online meetings. It was introduced as the “Personalized Assistant for Real-time Support in Explaining Jargon,” with a design centered on live transcript processing, user-specific filtering, and unobtrusive interface integration. In the broader literature, ParseJargon also sits within a research trajectory that treats jargon as reader-relative rather than intrinsic to a term, and that links jargon support to personalized familiarity modeling, definition generation, and de-jargonization workflows (Song et al., 13 Aug 2025, Guo et al., 2023, Wu et al., 22 May 2025, Huang et al., 2021, Nishal et al., 2024).
1. Problem setting and conceptual definition
In the ParseJargon system paper, jargon is defined as specialized terminology used within particular fields, including technical terms, acronyms, and discipline-specific phrases such as “Fourier neural operator,” “quasi-experimental design,” and “remote sensing.” The problem setting is interdisciplinary workplace communication, especially live online meetings in which participants from distinct professional backgrounds encounter domain-specific vocabulary under time pressure. The motivating claim is that speakers often assume shared knowledge that listeners do not in fact possess, which lowers comprehension, produces misunderstanding, reduces engagement, and can lead listeners to undervalue colleagues’ work (Song et al., 13 Aug 2025).
Related work on personalized jargon identification sharpens this definition by treating jargon as reader-relative. In that formulation, a term is jargon when its meaning and usage are not sufficiently familiar to a given reader, given their background. This view is operationalized through familiarity labels, where the same term may be familiar to one researcher and unfamiliar to another, even within the same broad discipline (Guo et al., 2023, Wu et al., 22 May 2025).
This reader-relative perspective is empirically important. One study reports over 10K term familiarity annotations from 11 computer science researchers on terms drawn from 100 paper abstracts and finds that only 19% of terms received uniform familiarity ratings from all 11 annotators, while around 15% of terms showed a split pattern in which roughly half of annotators rated them familiar and half unfamiliar (Guo et al., 2023). A later study on efficient personalization uses the same general framing and defines familiarity as a binary distinction between readers who “know this subject well and can describe it to others” and readers who “have encountered it but know little, or have never come across it at all” (Wu et al., 22 May 2025). This suggests that ParseJargon is best understood not merely as a glossary generator, but as a personalized inference system over term familiarity, explanation need, and interaction context.
2. Origins in online meeting research
The immediate design basis for ParseJargon is a two-week diary study with 16 professionals in a large tech company: 7 researchers, 5 engineers, 2 marketers, and 2 directors. Participants logged jargon encountered in 47 meetings, producing 123 terms, or about 2.6 terms per meeting. For each unfamiliar term they recorded the meeting context and one of six coping strategies: wait for explanation, skip the term, interrupt and ask, ask afterward, search internally, or search externally (Song et al., 13 Aug 2025).
The diary data establish a characteristic asymmetry between effectiveness and actual usage. Waiting for explanation was used for 56 of 123 terms, or 45.5%, with helpfulness approximately $3.11/5$, while interrupt-and-ask was used for 6 of 123 terms, or 4.9%, with helpfulness approximately $4.5$, and ask-afterward was used for 10 of 123 terms, or 8.1%, with helpfulness approximately $4.4$. External search was used for 31 of 123 terms, or 25.2%, with helpfulness approximately $3.52$, and internal search for 8 of 123 terms, or 6.5%, with helpfulness approximately $3.62$. Passive strategies were therefore most common and least effective; asking was most effective and least used; search was somewhat helpful but disruptive (Song et al., 13 Aug 2025).
The study also distinguishes same-domain from cross-domain jargon. Participants were more willing to ask about same-domain terms, about 22.2% of same-domain terms versus 3.6% of cross-domain terms, while passive strategies were more common for cross-domain terms, 60.7% versus 42.3%. Helpfulness ratings were lower for cross-domain jargon: waiting scored 3.5 for same-domain versus 2.97 for cross-domain, and searching scored 3.94 versus 3.20. These findings motivate four system requirements stated in the ParseJargon paper: real-time and automatic operation, personalization to each listener’s background, unobtrusive integration into existing workflows, and support for comprehension without increasing distraction (Song et al., 13 Aug 2025).
3. Architecture and real-time operation
ParseJargon is implemented as a Chrome extension front-end integrated with the Zoom web client, a Python Flask backend deployed on Heroku, and a PostgreSQL database for storing glossary terms and user feedback. The runtime inputs are live meeting transcripts from platform speech-to-text, a short textual user background profile, and user feedback or preferences. The outputs are detected jargon terms with plain-language definitions, a personalized glossary, highlighted terms in captions, and a sidebar glossary listing jargon encountered so far (Song et al., 13 Aug 2025).
The core inference pipeline is two-stage. Stage 1 performs jargon identification and explanation generation. GPT receives a transcript sentence, previously defined terms, and user preference information, and returns a JSON list of pairs. The prompt instructs the model to identify terms the audience might not understand, provide concise plain-language definitions, skip nonsense or transcription errors, and avoid repeating previously defined terms. Stage 2 performs personalized filtering. Given the raw glossary and the audience background, a second GPT call removes terms the user is likely already to understand and returns both understood terms and a refined glossary (Song et al., 13 Aug 2025).
The system therefore separates generic candidate detection from user-specific suppression. The paper states that this two-step design was empirically better than attempting to identify “personalized jargon” in a single prompt. The user model is lightweight: a free-text profile string, such as “I am a quantum computing researcher and hold a Physics PhD,” plus a history of likes and dislikes on prior explanations. A plausible implication is that personalization in the deployed system is primarily selection-oriented rather than representation-oriented: the model mostly decides what to surface, not yet how to deeply adapt explanatory style (Song et al., 13 Aug 2025).
Real-time operation is incremental. The extension intercepts live Zoom captions sentence by sentence and sends them to the backend. In the field deployment, GPT-4o-mini was used to keep latency low. The interface ensures that each new “latest term” is shown for at least 7 seconds, based on reading speed estimates of approximately 238 words per minute. The user interface comprises real-time captions with highlighted jargon, a “latest jargon definition” panel, and a persistent glossary sidebar (Song et al., 13 Aug 2025).
4. Personalization mechanisms
Personalization in ParseJargon has two explicit components. The first is term selection. In the paper’s example, for a talk combining deep learning and earth science, a machine learning engineer is expected to find “remote sensing” and “satellite data” unfamiliar while finding “foundation models” and “self-supervised learning” familiar, whereas an earth science researcher exhibits the reverse pattern. The second component is limited explanation framing: all explanations are concise and in plain language, but the current implementation does not yet deeply adapt explanation complexity or analogies to each user’s domain (Song et al., 13 Aug 2025).
This operational notion of personalization aligns with the earlier literature on personalized jargon identification. One study models familiarity using domain-level, subdomain-level, and individual-level features, including frequency, specificity, and embedding similarity derived from the reader’s own publications, and finds that abstract domain is universally important while individual-level frequency, specificity, and embedding similarity also frequently carry predictive weight (Guo et al., 2023). Another study investigates efficient personalization through Low-Rank Adaptation on Llama-3.1 8B Instruct and prompt-based personalization, using user metadata, profile summaries, nearest annotator signals, and nearest abstract signals as “Additional information” in the prompt (Wu et al., 22 May 2025).
The quantitative findings in that later study are directly relevant to the design space ParseJargon occupies. On the test set, GPT-4 prompt-based prediction achieved F1 64.2 ± 1.5, an oracle classifier achieved F1 71.9 ± 1.7, a personalized LoRA model with full training data achieved F1 77.9 ± 1.2, and a mixed regime with 10% supervised data plus unsupervised background signals achieved F1 77.0 ± 1.1 (Wu et al., 22 May 2025). The paper also reports that the personalized LoRA model outperforms GPT-4 by 21.4% in F1 score and exceeds the best performing oracle baseline by 8.3%, while using only 10% of the annotated training data can yield comparable performance in the mixed regime (Wu et al., 22 May 2025).
A plausible implication is that ParseJargon’s current free-text profile approach is one point in a broader methodological spectrum. Prompt-based personalization with user publications already has strong empirical support (Guo et al., 2023), while parameter-efficient personalization suggests a route toward persistent user-specific jargon models that remain computationally tractable (Wu et al., 22 May 2025).
5. Experimental results in meetings
The controlled evaluation of ParseJargon used a within-subjects design with 7 interns from the same company. Each participant watched six ten-minute project presentations, all prepared by the other interns, under three conditions: baseline (video plus transcript only), general-purpose ParseJargon (non-personalized glossary), and personalized ParseJargon (profile-filtered glossary). No pausing or scrubbing was allowed, and participants could use external tools if desired. Measures included self-reported comprehension confidence, engagement, perceived value of the work, glossary helpfulness, and presenter-rated quality of takeaways and questions (Song et al., 13 Aug 2025).
The quantitative results are strong for the personalized condition. Mean comprehension ratings were for baseline, for general-purpose support, and for personalized support. General versus baseline yielded , corrected $4.5$0, Cohen’s $4.5$1; personalized versus baseline yielded $4.5$2, $4.5$3, $4.5$4; personalized versus general yielded $4.5$5, $4.5$6, $4.5$7. Engagement was $4.5$8 for baseline, $4.5$9 for general, and $4.4$0 for personalized; personalized versus general gave $4.4$1, $4.4$2, $4.4$3. Perceived value was $4.4$4, $4.4$5, and $4.4$6, with personalized versus baseline giving $4.4$7, $4.4$8, $4.4$9 (Song et al., 13 Aug 2025).
Presenter-rated comprehension showed smaller effects on clarity but clear gains on relevance and depth. Relevance ratings were $3.52$0 for baseline, $3.52$1 for general, and $3.52$2 for personalized. Depth ratings were $3.52$3, $3.52$4, and $3.52$5, with personalized versus baseline giving $3.52$6, corrected $3.52$7, $3.52$8. The paper’s example from an ocean surface forecasting talk shows a progression from a generic takeaway in baseline, to a more relevant one in the general condition, to a technically deeper question in the personalized condition (Song et al., 13 Aug 2025).
The clearest mechanism-level result concerns overload. General-purpose support surfaced on average $3.52$9 terms per talk, of which $3.62$0 were marked helpful, for a helpfulness rate of $3.62$1. Personalized support surfaced $3.62$2 terms, of which $3.62$3 were marked helpful, for a helpfulness rate of $3.62$4, with the difference significant at $3.62$5. Qualitative responses describe the general condition as “too many term definitions with very short reading time,” “annoying,” and even “offensive” because it treated participants “like I know nothing,” whereas the personalized condition was experienced as more relevant and less disruptive (Song et al., 13 Aug 2025). This is consistent with the paper’s abstract claim that personalized jargon support significantly enhanced comprehension, engagement, and appreciation of colleagues’ work, whereas general-purpose support negatively affected engagement (Song et al., 13 Aug 2025).
A follow-up field study in a weekly Zoom meeting with 10 team members further evaluated real-time deployment. Usability ratings were $3.62$6 for ease of use, $3.62$7 for feature integration, and $3.62$8 for willingness for frequent future use. NASA-TLX scores were low: mental demand $3.62$9, physical demand 0, temporal demand 1, frustration 2, and effort 3. Participants reported that the system was automatic and non-intrusive and that it reduced the need to search externally during meetings, though two participants reported low effectiveness because of ASR errors or irrelevant jargon identification (Song et al., 13 Aug 2025).
6. Related methodological foundations
ParseJargon’s personalization layer is closely connected to the broader literature on reader-specific jargon modeling. “Personalized Jargon Identification for Enhanced Interdisciplinary Communication” develops familiarity prediction over 10,571 annotations from 11 computer science researchers and compares individual Lasso models, mixed Lasso models, and GPT-4 prompting with metadata, context-enhanced learning, and few-shot examples. The paper reports that prompt-based methods, especially those incorporating personal publications, yield the highest accuracy, while zero-shot prompting is already a strong baseline (Guo et al., 2023). This directly supports ParseJargon’s reliance on background-aware prompting rather than a purely generic glossary.
“Explain Less, Understand More: Jargon Detection via Personalized Parameter-Efficient Fine-tuning” moves from prompt-based personalization to scalable user adaptation with Low-Rank Adaptation. It defines four tasks—term familiarity plus needs for definition, background, and example—and shows that personalized LoRA on Llama-3.1 8B Instruct, 4-bit quantized, achieves F1 77.9 ± 1.2 on familiarity, while the mixed unsupervised plus 10% supervised regime achieves F1 77.0 ± 1.1 (Wu et al., 22 May 2025). The paper also reports that models trained only on familiarity labels generalize well to predicting definition need, background need, and example need. A plausible implication is that future versions of ParseJargon could replace or augment prompt-based filtering with user adapters that jointly model familiarity and explanation type.
The explanatory side of ParseJargon relates to two additional strands. “Understanding Jargon: Combining Extraction and Generation for Definition Modeling” proposes a combined extraction-and-generation framework for jargon definitions, using self-definitional information from the web, correlative definitional information from Wikipedia, and a BART generator conditioned on both. On UJ-CS, the combined model improves BLEU from 8.76 to 22.66 and human-annotated score from 2.34 to 4.04 (Huang et al., 2021). “De-jargonizing Science for Journalists with GPT-4: A Pilot Study” evaluates a human-in-the-loop system for scientific abstracts and finds median recall 0.68, median precision 0.33, and median 4 0.55 for jargon detection, while abstract-only definition generation yields 96.6% correct definitions versus 93.5% for RAG-based full-text context (Nishal et al., 2024). These results suggest that ParseJargon’s current concise-definition generation could, in principle, be strengthened by explicit definition extraction and evidence-grounded generation modules.
7. Limitations, design implications, and broader significance
The ParseJargon paper identifies several limitations. The controlled experiment used 7 participants, while the diary and field studies involved 16 and 10 participants respectively, all from one large tech company. Real-time performance depends on ASR quality, and the field deployment used GPT-4o-mini for latency, which was less capable than GPT-4o in the controlled study. Personalization currently relies on a short user-written profile and simple feedback history, so it may miss unreported expertise, and the system does not yet deeply personalize explanation complexity or modality (Song et al., 13 Aug 2025).
The related personalization literature adds further constraints. The 2023 familiarity study is restricted to computer science researchers reading non-CS abstracts, and the 2025 parameter-efficient study also uses a small annotator pool and reports instability for some unsupervised prompting strategies, including nearly 100% mismatch rates for nearest-annotator and nearest-abstract prompts in the unsupervised PEFT setting (Guo et al., 2023, Wu et al., 22 May 2025). The definition-modeling literature shows that jargon explanation is limited by retrieval quality and by residual hallucination in generation, even when extraction and generation are combined (Huang et al., 2021). The journalist-focused pilot likewise warns that over-flagging and low precision can clutter the interface, even when recall is high (Nishal et al., 2024).
Even with these constraints, the research record supports several durable design implications. First, personalization is essential: the controlled ParseJargon experiment shows that general-purpose support can raise comprehension while reducing engagement, whereas personalized support improves comprehension, engagement, and perceived value simultaneously (Song et al., 13 Aug 2025). Second, user modeling should be grounded in background corpora such as personal publications and subdomain exemplars rather than solely in self-description (Guo et al., 2023, Wu et al., 22 May 2025). Third, explanation generation should be selective and evidence-aware, because explanation overload and hallucinated downstream relevance are recurrent failure modes (Song et al., 13 Aug 2025, Huang et al., 2021).
In that sense, ParseJargon is both a specific system and a representative instantiation of a broader research program: personalized jargon support as an interaction problem, familiarity prediction as a reader-modeling problem, and definition generation as a grounded generation problem. The combined literature indicates that the most effective jargon support systems will likely couple real-time transcript or document analysis with personalized familiarity estimation, explicit prediction of information needs, and controlled explanation generation anchored in trusted sources (Song et al., 13 Aug 2025, Wu et al., 22 May 2025, Huang et al., 2021, Nishal et al., 2024).