Pith. sign in

REVIEW 3 major objections 4 minor 66 references

A single text box can route a smartphone user's raw typing to the correct app function: a week-long field study reports 71.35% top-1 accuracy, rising to 81.68% as the system learns each user.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

A unified text portal predicts users' intended text-related smartphone functions from raw input using a hybrid LLM+BERT model, reaching 71.35% Hit@1 in a real-world study.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection A genuinely new function-level portal idea with a sensible LLM+BERT hybrid, but the headline field accuracy likely overstates performance because the protocol let users test after the fact. the 3 major comments →

arxiv 2508.16926 v1 pith:RUAI65CL submitted 2025-08-23 cs.HC cs.AI

TextOnly: A Unified Function Portal for Text-Related Functions on Smartphones

Minghao Tu , Chun Yu , Xiyuan Shen , Zhi Zheng , Li Chen , Yuanchun Shi This is my paper
classification cs.HC cs.AI
keywords text-related functionsunified function portalintent predictionlarge language modelBERTpersonalizationsmartphone interactionservice retrieval
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper claims that the strings users type into any app—a search keyword, a message, a number to pay—carry enough signal to identify which text-related function the user wants, and that a single unified text box can therefore replace the multi-step navigation to the right text box. It builds TextOnly around two models: a large language model that supplies general knowledge and handles new users and new functions, and a BERT-based encoder that learns per-user preferences from accumulated history and produces fast local predictions. A confidence score routes each input either to the fast local integrator or to the slower LLM. In a week-long field study with 16 participants, top-1 accuracy reached 71.35% and rose from 46.51% on the first day to 81.68% by the end, while inference time fell from 2.73 seconds to 1.36 seconds as the BERT model took over most predictions. The result matters because text input is one of the most frequent phone actions, yet today each app requires finding its own text box manually.

Core claim

On the paper's own terms, the discovery is that brief raw text inputs—exactly the string a user would type into the destination text box—are predictive enough to act as a universal command line for text-related smartphone functions. The authors define functions as "application–action" pairs (e.g., Google-search, Memo-record, Facebook-chat), collect 5,136 real-world entries spanning 275 distinct functions, and show these concentrate in a power law while differing measurably between apps and between actions, which makes intent inferable from the input itself. The technical claim is that combining an LLM with a personalized BERT model beats either alone: the LLM solves the cold-start problem, t

What carries the argument

The system's load-bearing mechanism is the two-model inference loop. A BERT encoder converts the user's raw input plus app-usage and time context into a vector; the local integrator computes a similarity-weighted vote over the K=5 most similar historical inputs, with a user-weight coefficient (1.05) boosting same-user history. A confidence score—a rank-weighted average of those similarities—compared against a threshold of 0.95 decides whether to answer immediately with the local vote or query the LLM with a few-shot prompt built from the same similar instances. The LLM's ranked output is folded back, together with the user's actual selection, into a soft label that daily retrains the BERT mo

Load-bearing premise

The headline accuracy assumes the recorded trials reflect genuine spontaneous intentions; in the study, participants were required to use TextOnly rather than the app directly, the floating-ball reminder let them log a trial after the goal was already reached, and some deliberately simplified or altered their wording to get better predictions, so 71.35% likely overstates performance on unbiased first-time input.

What would settle it

Re-run the same prediction models on keystroke data captured invisibly from users' normal typing habits before any TextOnly exposure, then replay those exact strings through the system. If Hit@1 on these unadapted spontaneous inputs is close to the first-day level (46.51%) rather than the headline 71.35%, the personalization loop is real but the reported accuracy is inflated by user adaptation and post-hoc submission.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Users can reach a text-related function in one flow—type, tap the right suggestion, done—instead of opening an app and navigating to its text box.
  • Accuracy and speed self-improve with normal use: the system went from 46.51% to 81.68% Hit@1 in one week as the BERT model accumulated personal data.
  • The LLM-plus-small-model split is empirically validated: each alone is significantly worse than the pair, and the small model increasingly absorbs inference load (72.84% of predictions by day six).
  • Chat-oriented intents, excluded from the field study, are separately testable: intent classification hit 96.01% and contact prediction should improve through the same personalization loop.
  • A unified text portal covers a wider function range than built-in text portals (100% vs 52.59% of tested intentions) with shorter inputs than voice assistants (5.20 vs 11.69 words on average).

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The 71.35% figure measures a favorable setting: participants were required to use TextOnly, could submit trials after completing the task, and some deliberately simplified their phrasing (e.g., typing 'cold medication' instead of a full question). Spontaneous, unadapted input would likely score lower than the headline number, though the improvement-over-time trend may be unaffected.
  • The same hybrid pattern—a general LLM for cold start, a small trainable encoder for personalization, and confidence-based routing—is a template for other intent-prediction surfaces, such as clipboard text, shared links, or camera captures, where the raw artifact itself is the command.
  • Because every accepted prediction generates a labeled training example, TextOnly's data pipeline makes user behavior the curriculum: the system improves fastest for the functions users actually pick, a testable design principle for personalized retrieval beyond smartphones.
  • The requirement to record RPA scripts to add new functions is the main adoption bottleneck; automating function discovery from screen content, which the paper floats as future work, is the natural test of whether the portal generalizes beyond a fixed function set.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 4 minor

Summary. The paper introduces TextOnly, an Android 'unified function portal' that accepts a user's raw text and predicts the intended text-related function (app+action), then executes it via RPA. The system combines GPT-3.5 as a general-knowledge predictor with a per-user BERT encoder and a local nearest-neighbor integrator. The authors report a week-long field study (16 users, 3,847 trials) with Hit@1=0.7135, Hit@5=0.8994, MRR=0.7897, improving daily (46.51% to 81.68%) and becoming faster over time as BERT takes over; they also report a lab study for chat/contact prediction and comparisons with built-in portals and voice assistants. The paper claims that TextOnly outperforms baselines, that the LLM+BERT fusion helps, and that usability and satisfaction are high.

Significance. If the reported accuracy is valid, the contribution is significant: TextOnly offers a new interaction modality at function level, addresses cold-start through an LLM plus lightweight BERT personalization, and provides rare real-world evaluation data (5,136 collection entries plus 3,847 trials). The paper is transparent about prompt design and explicitly documents users' input adaptation. However, the central 'practicality' claim rests on a field measurement whose construct validity is weakened by post-hoc trials and user-adapted inputs; the current evidence is not yet sufficient to establish the headline numbers. Re-analysis is feasible and should be required.

major comments (3)
  1. [§5.1, §5.2.1, §5.2.4] The headline accuracy is vulnerable to a construct-validity bias that the paper itself documents. The protocol explicitly retained the floating-ball reminder so that participants could test TextOnly after completing an intention. Because Eq. (1) concatenates recent-app usage into the query vector, a post-hoc trial made shortly after the user has just used the target app leaks the target function into the context feature; Table 2 suggests context is worth ~9.5 Hit@1 points (0.6961 vs. 0.6011), so even a modest share of post-hoc trials can materially inflate the estimate. Independently, §5.2.4 reports that participants deliberately simplified or reformulated inputs to improve predictions, so the tested distribution is not spontaneous natural language. The paper does not report a live/post-hoc split or an original-vs-adapted analysis; without this, the reported 71.35% Hit@1 cannot be read a
  2. [§3.3–§3.5] The framework depends on hyperparameters (T, K, C_user_weight, M, label-weight distribution, α) reported as fixed 'after experiments'/'after testing' with no sensitivity analysis or validation split. If these values were chosen using the same real-world study data, the reported accuracy and ablation comparisons are optimistically biased. Please provide a description of when and how these values were selected, and add robustness analyses (leave-one-participant-out or sensitivity sweeps) for at least T, K, and C_user_weight.
  3. [§5.2.2 and Table 2] The claim that TextOnly 'significantly' outperforms all compared models is stronger than the reported tests support. On Hit@5, TextOnly (0.8660) is not marked significant over LLM-only (0.8294) or TextOnly-nocontext (0.8375); on MRR, it is not marked significant over TextOnly-nocontext. The narrative should qualify the significance statements and, if possible, report paired effect sizes and confidence intervals.
minor comments (4)
  1. [§5.2.2] The paragraph opens with 'Table 3 presents...' but the referenced table is Table 2.
  2. [§4.2.3, §5.1, §5.2.2] Typos: 'ont-hot' should be 'one-hot' (§4.2.3); 'candidiates' should be 'candidates' (§5.1); 'neccesity' should be 'necessity' (§5.2.2).
  3. [§4.1 and §5.1] Please state explicitly whether the 16 user-study participants overlap with the 22 data-collection participants; this affects how the pre-training data are interpreted.
  4. [Appendix/References] The 'Received 20 February 2007' date at the end appears to be a template artifact and should be corrected.

Circularity Check

0 steps flagged

No significant circularity: the central claims are empirical measurements, not derivations that reduce to their inputs.

full rationale

The paper's central claims — TextOnly's Hit@1 of 0.7135, Hit@5 of 0.8994, and its improvements over time — are empirical results from a real-world user study (Section 5.2.1, Figure 7), not the outcome of a fitted parameter being renamed as a prediction. The BERT model is trained partly on the LLM's top-5 outputs to form soft labels (Section 3.5), creating a self-training feedback loop, but this is not circular reasoning about the measured accuracy: the headline metric is still evaluated against users' actual selections, and the ablation comparisons (TextOnly vs. BERT-only vs. LLM-only in Table 2) are empirical comparisons rather than constructed equivalences. The only prominent self-citation, future scanner [23] for RPA execution and triggering (Section 3.1), is an implementation dependency and is not load-bearing for the accuracy claims. The potential threats from post-hoc trials and user-adapted inputs (Sections 5.1 and 5.2.4) are validity concerns, not circularity: they do not make the reported accuracy equal to the model's own inputs by construction. No equation defines the target quantity in terms of the system's predictions, and no uniqueness theorem or ansatz is imported from prior work to force the model choice. Therefore, no circular step is present.

Axiom & Free-Parameter Ledger

6 free parameters · 4 axioms · 0 invented entities

No new physical or formal entities are introduced. The system combines existing tools: GPT-3.5, BERT, and the authors' own Scanner RPA app. The free parameters above are tuned on in-house data and directly affect the reported accuracy and speed.

free parameters (6)
  • Threshold T for local integrator takeover = 0.95
    Set after experiments (Section 3.4); controls how often LLM is queried, directly affects accuracy and speed.
  • K (number of similar history instances) = 5
    Fixed after testing (Section 3.4); used in similarity search and local integrator.
  • C_user_weight = 1.05
    Coefficient boosting same-user similarity in Equation 3; tuned after experiments (Section 3.4).
  • Label weight distribution = [0.8, 0.07, 0.06, 0.04, 0.03]
    Weights for combining user selection with LLM top-5 output when forming BERT training labels (Section 3.5); adopted after testing.
  • M (few-shot examples count) = 20
    Number of similar historical instances placed in LLM prompt; selected after experiments (Section 3.3).
  • alpha (pre-training data scaling) = 10
    Scales the amount of initial personal data selected for each user (Section 3.5).
axioms (4)
  • domain assumption User raw text inputs carry enough signal to distinguish intended text-related functions.
    Central premise: Section 1 and Section 4.2.2 argue inputs differ across apps/actions (e.g., addresses vs food names), enabling prediction. If inputs were not discriminative, the whole approach fails.
  • domain assumption The population and function distribution in the studies are representative of general smartphone users.
    Section 4.2.1: 22 student/staff participants, search dominates (top 20 = 76.35%). Accuracy may not generalize to users with a different function mix, especially chat-oriented functions which were excluded from data collection.
  • domain assumption GPT-3.5 provides sufficiently accurate general-knowledge inference for function ranking.
    Section 3.3: the LLM is the principal inference component, unfine-tuned, queried via API. The system inherits its ceiling (also acknowledged in Section 6).
  • domain assumption A BERT encoder retrained on personal data and LLM outputs can capture user preferences without catastrophic interference.
    Section 3.4-3.5: daily retraining on a small personal database; the claimed improvement over time depends on this assumption.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of TextOnly: A Unified Function Portal for Text-Related Functions on Smartphones." pith.science (2026). https://pith.science/paper/RUAI65CL

@misc{pith2026250816926,
  author       = {Pith},
  title        = {Pith review of: TextOnly: A Unified Function Portal for Text-Related Functions on Smartphones},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RUAI65CL}},
  note         = {Machine review of arXiv:2508.16926}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Text boxes serve as portals to diverse functionalities in today's smartphone applications. However, when it comes to specific functionalities, users always need to navigate through multiple steps to access particular text boxes for input. We propose TextOnly, a unified function portal that enables users to access text-related functions from various applications by simply inputting text into a sole text box. For instance, entering a restaurant name could trigger a Google Maps search, while a greeting could initiate a conversation in WhatsApp. Despite their brevity, TextOnly maximizes the utilization of these raw text inputs, which contain rich information, to interpret user intentions effectively. TextOnly integrates large language models(LLM) and a BERT model. The LLM consistently provides general knowledge, while the BERT model can continuously learn user-specific preferences and enable quicker predictions. Real-world user studies demonstrated TextOnly's effectiveness with a top-1 accuracy of 71.35%, and its ability to continuously improve both its accuracy and inference speed. Participants perceived TextOnly as having satisfactory usability and expressed a preference for TextOnly over manual executions. Compared with voice assistants, TextOnly supports a greater range of text-related functions and allows for more concise inputs.

Figures

Figures reproduced from arXiv: 2508.16926 by Chun Yu, Li Chen, Minghao Tu, Xiyuan Shen, Yuanchun Shi, Zhi Zheng.

Figure 1
Figure 1. Figure 1: TextOnly redirects users’ input to their intended text boxes. The input of TextOnly is exactly what the user types into his intended text box. the workflow of function navigation, such as app selection via text query[2, 3] and app recommendation that predicts the next apps to be used[25, 35, 36]. While these solutions help users access their desired apps, they fall short in seamlessly guiding users to thei… view at source ↗
Figure 2
Figure 2. Figure 2: Interfaces of TextOnly. Figure (a) illustrates the initiation of TextOnly from the home screen; Figure (b) is the input interface of TextOnly; Figure (c) is the function selection interface of TextOnly, where the user’s input is “100”. users can trigger TextOnly by clicking on the floating ball on the screen. TextOnly can also be triggered by the user’s tap-tap gesture, which is supported by future scanner… view at source ↗
Figure 3
Figure 3. Figure 3: Model framework of TextOnly Due to privacy concerns, we excluded chat-oriented intentions during data collection. However, in the imple￾mentation of TextOnly, we still offer support for these chat-oriented intentions. For such intentions, we need to further specify the contact to chat with. Therefore, we represent chatting functions in the form of “app-contact”. Once a chatting function is selected, TextOn… view at source ↗
Figure 4
Figure 4. Figure 4: Number of data entries for top 20 text-related functions. [PITH_FULL_IMAGE:figures/full_fig_p010_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Average length of input texts of each action [PITH_FULL_IMAGE:figures/full_fig_p011_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Similarity heatmaps of text-related functions between participants [PITH_FULL_IMAGE:figures/full_fig_p012_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Daily performance of TextOnly during user study [PITH_FULL_IMAGE:figures/full_fig_p015_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Comparison of satisfaction score between TextOnly and built-in text portal [PITH_FULL_IMAGE:figures/full_fig_p018_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Examples of supported functions and unsupported functions of TextOnly. The text box in (a) is fixed and the text [PITH_FULL_IMAGE:figures/full_fig_p020_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: An example of the prompt for function prediction. In real usage, there are more use cases. [PITH_FULL_IMAGE:figures/full_fig_p026_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: An example of the prompt for contact selection. In real usage, chat histories with all contacts are included. [PITH_FULL_IMAGE:figures/full_fig_p027_11.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

66 extracted references · 53 canonical work pages

  1. [1]

    42matters. 2024. Google Play Statistics and Trends 2024 . https://42matters.com/google-play-statistics-and-trends/

    work page 2024

  2. [2]

    Mohammad Aliannejadi, Hamed Zamani, Fabio Crestani, and W Bruce Croft. 2018. Target apps selection: Towards a unified search framework for mobile devices. In The 41st international acm sigir conference on research & development in information retrieval . 215–224

    work page 2018

  3. [3]

    Mohammad Aliannejadi, Hamed Zamani, Fabio Crestani, and W Bruce Croft. 2021. Context-aware target apps selection and recommen- dation for enhancing personal mobile assistants. ACM Transactions on Information Systems (TOIS) 39, 3 (2021), 1–30

    work page 2021

  4. [4]

    Amos Azaria, Jayant Krishnamurthy, and Tom Mitchell. 2016. Instructable intelligent personal agent. In Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 30

    work page 2016

  5. [5]

    Ricardo Baeza-Yates, Di Jiang, Fabrizio Silvestri, and Beverly Harrison. 2015. Predicting the next app that you are going to use. In Proceedings of the eighth ACM international conference on web search and data mining . 285–294

    work page 2015

  6. [6]

    Jonathan Berant, Andrew Chou, Roy Frostig, and Percy Liang. 2013. Semantic parsing on freebase from question-answer pairs. In Proceedings of the 2013 conference on empirical methods in natural language processing . 1533–1544

    2013

  7. [7]

    Rishi Bommasani, Drew A Hudson, Ehsan Adeli, Russ Altman, Simran Arora, Sydney von Arx, Michael S Bernstein, Jeannette Bohg, Antoine Bosselut, Emma Brunskill, et al. 2021. On the opportunities and risks of foundation models. arXiv preprint arXiv:2108.07258 (2021)

    Pith/arXiv arXiv 2021

  8. [8]

    Michelle Brachman, Christopher Bygrave, Tathagata Chakraborti, Arunima Chaudhary, Zhining Ding, Casey Dugan, David Gros, Thomas Gschwind, James Johnson, Jim Laredo, et al. 2022. A Goal-driven natural language interface for creating application integration workflows. In Proceedings of the AAAI Conference on Artificial Intelligence , Vol. 36. 13155–13157

    work page 2022

  9. [9]

    Tom Broens, Stanislav Pokraev, Marten Van Sinderen, Johan Koolwaaij, and Patricia Dockhorn Costa. 2004. Context-aware, ontology- based service discovery. In Ambient Intelligence: Second European Symposium, EUSAI 2004, Eindhoven, The Netherlands, November 8-11,

    work page 2004

  10. [10]

    Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. 2020. Language models are few-shot learners. Advances in neural information processing systems 33 (2020), 1877–1901

    2020

  11. [11]

    Buqing Cao, Jianxun Liu, Yiping Wen, Hongtao Li, Qiaoxiang Xiao, and Jinjun Chen. 2019. QoS-aware service recommendation based on relational topic model and factorization machines for IoT Mashup applications. Journal of parallel and distributed computing 132 (2019), 177–189. Proc. ACM Meas. Anal. Comput. Syst., Vol. 37, No. 4, Article 111. Publication dat...

    work page 2019

  12. [12]

    Gilbert Cassar, Payam Barnaghi, Wei Wang, and Klaus Moessner. 2012. A hybrid semantic matchmaker for IoT services. In2012 IEEE International Conference on Green Computing and Communications . IEEE, 210–216

    work page 2012

  13. [13]

    Cheng Chen, Takuya Maekawa, Daichi Amagata, and Takahiro Hara. 2021. Predicting Next-use Mobile Apps Using App Semantic Representations. Journal of Information Processing 29 (2021), 597–609

    work page 2021

  14. [14]

    Lingjiao Chen, Matei Zaharia, and James Zou. 2023. FrugalGPT: How to Use Large Language Models While Reducing Cost and Improving Performance. arXiv preprint arXiv:2305.05176 (2023)

    Pith/arXiv arXiv 2023

  15. [15]

    Weihao Chen, Chun Yu, Huadong Wang, Zheng Wang, Lichen Yang, Yukun Wang, Weinan Shi, and Yuanchun Shi. 2023. From Gap to Synergy: Enhancing Contextual Understanding through Human-Machine Collaboration in Personalized Systems. In Proceedings of the 36th Annual ACM Symposium on User Interface Software and Technology . 1–15

    work page 2023

  16. [16]

    Xinlei Chen, Yu Wang, Jiayou He, Shijia Pan, Yong Li, and Pei Zhang. 2019. CAP: Context-aware app usage prediction with heterogeneous graph embedding. Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies 3, 1 (2019), 1–25

    work page 2019

  17. [17]

    Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, et al. 2022. Palm: Scaling language modeling with pathways. arXiv preprint arXiv:2204.02311 (2022)

    Pith/arXiv arXiv 2022

  18. [18]

    Karen Church, Barry Smyth, Keith Bradley, and Paul Cotter. 2008. A large scale study of European mobile search behaviour. InProceedings of the 10th international conference on Human computer interaction with mobile devices and services . 13–22

    work page 2008

  19. [19]

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2018. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805 (2018)

    Pith/arXiv arXiv 2018

  20. [20]

    Google. 2024. Android Accessibility Service. https://developer.android.com/reference/android/accessibilityservice/AccessibilityService

    work page 2024

  21. [21]

    Izzeddin Gur, Ulrich Rueckert, Aleksandra Faust, and Dilek Hakkani-Tur. 2018. Learning to navigate the web. arXiv preprint arXiv:1812.09195 (2018)

    Pith/arXiv arXiv 2018

  22. [22]

    Ido Guy. 2016. Searching by talking: Analysis of voice queries on mobile web search. In Proceedings of the 39th International ACM SIGIR conference on Research and Development in Information Retrieval . 35–44

    work page 2016

  23. [23]

    Hcifuture. 2023. Scanner. https://app.mi.com/details?id=com.hcifuture.scanner&ref=search

    work page 2023

  24. [24]

    Peter Hofmann, Caroline Samp, and Nils Urbach. 2020. Robotic process automation. Electronic markets 30, 1 (2020), 99–106

    work page 2020

  25. [25]

    Ke Huang, Chunhui Zhang, Xiaoxiao Ma, and Guanling Chen. 2012. Predicting mobile application usage using contextual information. In proceedings of the 2012 ACM conference on ubiquitous computing . 1059–1065

    work page 2012

  26. [26]

    Peter C Humphreys, David Raposo, Tobias Pohlen, Gregory Thornton, Rachita Chhaparia, Alistair Muldal, Josh Abramson, Petko Georgiev, Adam Santoro, and Timothy Lillicrap. 2022. A data-driven approach for learning to control computers. In International Conference on Machine Learning . PMLR, 9466–9482

    work page 2022

  27. [27]

    Maryam Kamvar and Shumeet Baluja. 2006. A large scale study of wireless search behavior: Google mobile search. In Proceedings of the SIGCHI conference on Human Factors in computing systems . 701–709

    work page 2006

  28. [28]

    Yonchanok Khaokaew, Hao Xue, and Flora D Salim. 2024. MAPLE: Mobile App Prediction Leveraging Large Language model Embeddings. Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies 8, 1 (2024), 1–25

    work page 2024

  29. [29]

    Matthias Klusch, Benedikt Fries, and Katia Sycara. 2006. Automated semantic web service discovery with OWLS-MX. In Proceedings of the fifth international joint conference on Autonomous agents and multiagent systems . 915–922

    work page 2006

  30. [30]

    Matthias Klusch and Patrick Kapahnke. 2008. Semantic web service selection with SAWSDL-MX. In The 7th International Semantic Web Conference. Citeseer, 3

    work page 2008

  31. [31]

    Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. 2022. Large language models are zero-shot reasoners. arXiv preprint arXiv:2205.11916 (2022)

    Pith/arXiv arXiv 2022

  32. [32]

    Tessa Lau, Julian Cerruti, Guillermo Manzato, Mateo Bengualid, Jeffrey P Bigham, and Jeffrey Nichols. 2010. A conversational interface to web automation. In Proceedings of the 23nd annual ACM symposium on User interface software and technology . 229–238

    work page 2010

  33. [33]

    Yang Li, Jiacong He, Xin Zhou, Yuan Zhang, and Jason Baldridge. 2020. Mapping natural language instructions to mobile UI action sequences. arXiv preprint arXiv:2005.03776 (2020)

    Pith/arXiv arXiv 2020

  34. [34]

    Yuanchun Li, Baoxiong Jia, Yao Guo, and Xiangqun Chen. 2017. Mining user reviews for mobile app comparisons. Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies 1, 3 (2017), 1–15

    work page 2017

  35. [35]

    Zhung-Xun Liao, Po-Ruey Lei, Tsu-Jou Shen, Shou-Chung Li, and Wen-Chih Peng. 2012. Mining temporal profiles of mobile applications for usage prediction. In 2012 IEEE 12th International Conference on Data Mining Workshops . IEEE, 890–893

    work page 2012

  36. [36]

    Eric Hsueh-Chan Lu, Yi-Wei Lin, and Jing-Bin Ciou. 2014. Mining mobile application sequential patterns for usage prediction. In 2014 IEEE International Conference on Granular Computing (GrC) . IEEE, 185–190

    work page 2014

  37. [37]

    Like Having a Really Bad PA

    Ewa Luger and Abigail Sellen. 2016. " Like Having a Really Bad PA" The Gulf between User Expectation and Experience of Conversational Agents. In Proceedings of the 2016 CHI conference on human factors in computing systems . 5286–5297

    work page 2016

  38. [38]

    Jiangang Ma, Yanchun Zhang, and Jing He. 2008. Efficiently finding web services using a clustering semantic approach. In Proceedings of the 2008 international workshop on Context enabled source and service selection, integration and adaptation: organized with the 17th International World Wide Web Conference (WWW 2008) . 1–8. Proc. ACM Meas. Anal. Comput. ...

    work page 2008

  39. [39]

    Georgios Meditskos and Nick Bassiliades. 2009. Structural and role-oriented web service discovery with taxonomies in OWL-S. IEEE transactions on knowledge and data engineering 22, 2 (2009), 278–290

    work page 2009

  40. [40]

    Chelsea Myers, Anushay Furqan, Jessica Nebolsky, Karina Caro, and Jichen Zhu. 2018. Patterns for how users overcome obstacles in voice user interfaces. In Proceedings of the 2018 CHI conference on human factors in computing systems . 1–7

    work page 2018

  41. [41]

    Aaron van den Oord, Yazhe Li, and Oriol Vinyals. 2018. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748 (2018)

    Pith/arXiv arXiv 2018

  42. [42]

    OpenAI. 2022. GPT-3.5 Turbo

    work page 2022

  43. [43]

    OpenAI. 2022. Introducing ChatGPT. https://openai.com/blog/chatgpt

    work page 2022

  44. [44]

    Lihang Pan, Chun Yu, JiaHui Li, Tian Huang, Xiaojun Bi, and Yuanchun Shi. 2022. Automatically generating and improving voice command interface from operation sequences on smartphones. In Proceedings of the 2022 CHI Conference on Human Factors in Computing Systems. 1–21

    work page 2022

  45. [45]

    Dae Hoon Park, Yi Fang, Mengwen Liu, and ChengXiang Zhai. 2016. Mobile app retrieval for social media users via inference of implicit intent in social media text. In Proceedings of the 25th ACM International on Conference on Information and Knowledge Management . 959–968

    work page 2016

  46. [46]

    Jyotishman Pathak, Neeraj Koul, Doina Caragea, and Vasant G Honavar. 2005. A framework for semantic web services discovery. In Proceedings of the 7th annual ACM international workshop on Web information and data management . 45–50

    work page 2005

  47. [47]

    O’Reilly Media, Inc

    Cathy Pearl. 2016. Designing voice user interfaces: Principles of conversational experiences . " O’Reilly Media, Inc. "

    work page 2016

  48. [48]

    Lenin Ravindranath, Arvind Thiagarajan, Hari Balakrishnan, and Samuel Madden. 2012. Code in the air: simplifying sensing and coordination tasks on smartphones. In Proceedings of the twelfth workshop on mobile computing systems & applications . 1–6

    work page 2012

  49. [49]

    Sesame. 2022. Sesame Shortcuts. http://sesame.ninja/

    work page 2022

  50. [50]

    Ben Shneiderman. 2000. The limits of speech recognition. Commun. ACM 43, 9 (2000), 63–65

    work page 2000

  51. [51]

    Milad Shokouhi, Rosie Jones, Umut Ozertem, Karthik Raghunathan, and Fernando Diaz. 2014. Mobile query reformulations. InProceedings of the 37th international ACM SIGIR conference on Research & development in information retrieval . 1011–1014

    work page 2014

  52. [52]

    Yu Su, Ahmed Hassan Awadallah, Madian Khabsa, Patrick Pantel, Michael Gamon, and Mark Encarnacion. 2017. Building natural language interfaces to web apis. In Proceedings of the 2017 ACM on Conference on Information and Knowledge Management . 177–186

    work page 2017

  53. [53]

    Craig W Thompson and Kenneth M Ross. 1987. Natural-language interface generating system. US Patent 4,688,195

    work page 1987

  54. [54]

    Eran Toch, Avigdor Gal, Iris Reinhartz-Berger, and Dov Dori. 2007. A semantic approach to approximate service retrieval. ACM Transactions on Internet Technology (TOIT) 8, 1 (2007), 2–es

    work page 2007

  55. [55]

    Huandong Wang, Yong Li, Mu Du, Zhenhui Li, and Depeng Jin. 2021. App2Vec: Context-aware application usage prediction. ACM Transactions on Knowledge Discovery from Data (TKDD) 15, 6 (2021), 1–21

    work page 2021

  56. [56]

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Ed Chi, Quoc Le, and Denny Zhou. 2022. Chain of thought prompting elicits reasoning in large language models. arXiv preprint arXiv:2201.11903 (2022)

    Pith/arXiv arXiv 2022

  57. [57]

    wikipedia. 2023. Spotlight (Apple). https://en.wikipedia.org/wiki/Spotlight_(Apple)

    work page 2023

  58. [58]

    Chenfei Wu, Shengming Yin, Weizhen Qi, Xiaodong Wang, Zecheng Tang, and Nan Duan. 2023. Visual chatgpt: Talking, drawing and editing with visual foundation models. arXiv preprint arXiv:2303.04671 (2023)

    Pith/arXiv arXiv 2023

  59. [59]

    Tong Xia, Yong Li, Jie Feng, Depeng Jin, Qing Zhang, Hengliang Luo, and Qingmin Liao. 2020. DeepApp: Predicting personalized smartphone app usage via context-aware multi-task learning. ACM Transactions on Intelligent Systems and Technology (TIST) 11, 6 (2020), 1–12

    work page 2020

  60. [60]

    Scott Wen-tau Yih, Ming-Wei Chang, Xiaodong He, and Jianfeng Gao. 2015. Semantic parsing via staged query graph generation: Question answering with knowledge base. In Proceedings of the Joint Conference of the 53rd Annual Meeting of the ACL and the 7th International Joint Conference on Natural Language Processing of the AFNLP

    work page 2015

  61. [61]

    Donghan Yu, Yong Li, Fengli Xu, Pengyu Zhang, and Vassilis Kostakos. 2018. Smartphone app usage prediction using points of interest. Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies 1, 4 (2018), 1–21

    work page 2018

  62. [62]

    Luke S Zettlemoyer and Michael Collins. 2012. Learning to map sentences to logical form: Structured classification with probabilistic categorial grammars. arXiv preprint arXiv:1207.1420 (2012)

    Pith/arXiv arXiv 2012

  63. [63]

    Feng Zhao, Zheng Sun, and Hai Jin. 2015. Topic-centric and semantic-aware retrieval system for internet of things. Information Fusion 23 (2015), 33–42

    work page 2015

  64. [64]

    Lina Zhou. 2007. Natural language interface for information management on mobile devices. Behaviour & Information Technology 26, 3 (2007), 197–207

    work page 2007

  65. [65]

    Lina Zhou, Ammar S Mohammed, and Dongsong Zhang. 2012. Mobile personal information management agent: Supporting natural language interface and application integration. Information Processing & Management 48, 1 (2012), 23–31. Proc. ACM Meas. Anal. Comput. Syst., Vol. 37, No. 4, Article 111. Publication date: August 2025. 111:24 • Tu, et al. A DETAILED INFO...

    work page 2012

  66. [2004]

    Springer, 72–83

    Proceedings 2. Springer, 72–83

    work page

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.