← Research

Research story · Text mining and topic modeling

What can statistics learn from what people write?

Reviews, comments, course evaluations and conference abstracts are written for people. This line of work turns them into data along several parallel threads, which grew from different problems: scheduling a large conference, reading opinions and emotions, building quality indexes and estimating what guests do not score.

At a glance

Period
2019 to the present
Threads
Conference scheduling; opinions and emotions; indexes from comments; reviews and places; topic extraction
Methods
Topic models, sentiment analysis, partial least squares path modeling, small area estimation
Selected collaborators
Marco Ortu, Giulia Contu, Andrea Carta, Maurizio Romano, Qiuyi Wu, David Banks. All authors are listed with each publication.
Key paper
Journal of the American Statistical Association 2022

The question

Reviews, comments and abstracts are written for people, who read them one at a time. Read by the thousands, they contain information that no single reader can see: the topics that recur, the mood in which they are discussed, the problems that guests or students mention but do not score. Text mining turns this unstructured material into data that statistical models can use.

1 · Texts and their words

Text 1 Text 2 Text 3 Text 4

2 · Each text as a mix of topics

Schematic illustration with invented texts: a topic model finds groups of words that tend to occur together, the topics (one colour each; grey words are not typical of any topic), and describes each text by the share of each topic.

A conference program that schedules itself 2022–2024

At large conferences such as the Joint Statistical Meetings (JSM), sessions on the same topic often run at the same time, and participants have to choose. We used seeded latent Dirichlet allocation, an existing topic model, together with a scheduling optimization algorithm to reduce the overlap of content between parallel sessions of the 2020 program (Journal of the American Statistical Association, 2022).

Topics collide in the same time slot

Each time slot offers different topics

Schematic illustration with simulated data, not the actual JSM 2020 schedule: five parallel rooms (columns) and five time slots (rows); colours are topics.

0.058 original JSM 2020 program

0.371 optimized schedule

Score based on total variation distance, from 0 (random scheduling) to 1 (no overlapping content). Journal of the American Statistical Association, 2022.

According to the article, an optimized schedule could also replace the traditional in-person meeting in which dozens of program chairs build the JSM program, a significant saving for the American Statistical Association. I presented the approach in an invited session of the Italian Statistical Society (SIS 2022) and, with work on the efficiency of time slots, at a conference on optimization and decision science (ODS 2024).

Opinions and emotions in reviews and social media 2019–2023

A study of Booking.com reviews of 619 accommodation facilities in Sardinia built a classifier, an adapted Naive Bayes model, that told positive from negative comments with an accuracy of 91%, the highest among the classifiers compared. The model was then used to identify the strengths and weaknesses of accommodation and of its destination and, through a random forest, to predict the score of a review (GARR conference, 2019).

In marketing research, a mixed-method study based on text mining examined value co-destruction in social media interactions (2019). The journal article of this thread studied problematic interactions between customers and firms on social media: it categorized customers’ comments as positive, neutral or negative, located the peaks of negative comments, classified problematic interactions and correlated customers’ emotions. The same pattern of emotions appeared in different interactions, with different intensities, and value co-creation and co-destruction occurred together (Journal of Research in Interactive Marketing, 2023).

A chapter in a handbook of interactive marketing followed the negative emotions in these interactions: it detected the negative comments, located the periods with peaks of them, grouped those periods by topic and characterized each topic by its emotions. It concludes that, on social media, value co-destruction can be understood as an exchange of negative comments (Palgrave, 2023).

Semi-supervised sentiment clustering, which groups reviews by their overall sentiment, is told in the story on clustering.

From students’ comments to quality indexes 2022–2024

A conference paper used natural language processing to explain how students assess their satisfaction (IES 2022). TOBIAS, which we developed, combines sentiment analysis and topic modeling with partial least squares path modeling to estimate how the topics, sentiments, emotions and moods of comments affect an overall rating. Applied to the evaluations of courses by students of the University of Cagliari, it gave interpretable results on how the topics discussed, and the feelings expressed, relate to the overall rating (Computational Statistics, 2024).

What guests do not score 2024–2025

Reviews of Airbnb accommodation tend to be more positive than those on other platforms, so many negative experiences remain untracked. We combined sentiment analysis and topic modeling with small area estimation, through a mixed effects random forest, to measure negative sentiment and critical topics in small areas of Rome, defined by geography and by the presence of hosts and Superhosts (Statistical Methods and Applications, 2024).

Two conference papers followed the same data source in other directions: Airbnb reviews used to explore spatial disparities in transportation and mobility in Sardinia (IES 2025), and sustainability-oriented self-descriptions mapped to Airbnb reviews (IES 2025).

Extracting topics 2023–2026

Two contributions work on topic extraction itself: multivariate regression trees for topic modeling (CLADAG 2023) and precision-oriented semantic filtering for extracting latent topics from multilingual texts (IFCS 2026).

Publications in this line

2026

  • Precision-Oriented Semantic Filtering for Latent Topic Extraction in Multilingual Textual Corpora

    M. Ortu, L. Frigau, G. Contu, A. Carta

    Book of Abstracts of IFCS 2026

    Cite
    @inproceedings{ortu2026precision,
  author    = {Ortu, M. and Frigau, L. and Contu, G. and Carta, A.},
  title     = {Precision-Oriented Semantic Filtering for Latent Topic Extraction in Multilingual Textual Corpora},
  booktitle = {Book of Abstracts of IFCS 2026},
  year      = {2026},
}

2025

  • Exploring Spatial Disparities and Gaps in Transportation and Mobility through Airbnb Reviews: A Sardinian Case Study

    L. Frigau, G. Contu, M. Ortu, F. Mola

    IES 2025 - Book of Short Papers, pp. 410–417

    Cite
    @inproceedings{frigau2025exploring,
  author    = {Frigau, L. and Contu, G. and Ortu, M. and Mola, F.},
  title     = {Exploring Spatial Disparities and Gaps in Transportation and Mobility through Airbnb Reviews: A Sardinian Case Study},
  booktitle = {IES 2025 - Book of Short Papers},
  year      = {2025},
  pages     = {410--417},
  isbn      = {978-88-5495-849-4},
}

  • From Self-Perception to Feedback: Mapping Sustainability-Oriented Self-Descriptions to Airbnb Reviews

    M. Ortu, G. Contu, L. Frigau

    IES 2025 - Book of Short Papers, pp. 1349–1355

    Cite
    @inproceedings{ortu2025self,
  author    = {Ortu, M. and Contu, G. and Frigau, L.},
  title     = {From Self-Perception to Feedback: Mapping Sustainability-Oriented Self-Descriptions to Airbnb Reviews},
  booktitle = {IES 2025 - Book of Short Papers},
  year      = {2025},
  pages     = {1349--1355},
  isbn      = {978-88-5495-849-4},
}

2024

  • Gauging Airbnb Review Sentiments and Critical Key-Topics by Small Area Estimation

    L. Frigau, G. Contu, M. Ortu, A. Carta

    Statistical Methods and Applications, 33(4): 1145–1170

    Small area estimation of negative sentiment and critical topics in Airbnb reviews, for areas of the city of Rome.

    doi:10.1007/s10260-024-00764-y
    Cite
    @article{frigau2024gauging,
  author    = {Frigau, L. and Contu, G. and Ortu, M. and Carta, A.},
  title     = {Gauging Airbnb Review Sentiments and Critical Key-Topics by Small Area Estimation},
  journal   = {Statistical Methods and Applications},
  year      = {2024},
  volume    = {33},
  number    = {4},
  pages     = {1145--1170},
  doi       = {10.1007/s10260-024-00764-y},
  issn      = {1613-981X},
}

  • Topic Based Quality Indexes Assessment Through Sentiment

    M. Ortu, L. Frigau, G. Contu

    Computational Statistics, 39(1): 289–311

    TOBIAS relates the topics and sentiments of students’ comments to the overall rating of university courses through partial least squares path modeling.

    doi:10.1007/s00180-022-01284-7
    Cite
    @article{ortu2024topic,
  author    = {Ortu, M. and Frigau, L. and Contu, G.},
  title     = {Topic Based Quality Indexes Assessment Through Sentiment},
  journal   = {Computational Statistics},
  year      = {2024},
  volume    = {39},
  number    = {1},
  pages     = {289--311},
  doi       = {10.1007/s00180-022-01284-7},
  issn      = {1613-9658},
}

  • Enhancing Time Slot Efficiency in Scientific Conferences by Employing Latent Dirichlet Allocation

    L. Frigau

    ODS 2024 - Book of Abstracts, p. 151

    Cite
    @inproceedings{frigau2024enhancing,
  author    = {Frigau, L.},
  title     = {Enhancing Time Slot Efficiency in Scientific Conferences by Employing Latent Dirichlet Allocation},
  booktitle = {ODS 2024 - Book of Abstracts},
  year      = {2024},
  pages     = {151},
}

2023

  • How emotions impact the interactive value formation process during problematic social media interactions

    M. Frau, F. Cabiddu, L. Frigau, P. Tomczyk, F. Mola

    Journal of Research in Interactive Marketing, 17(5): 773–793

    doi:10.1108/JRIM-06-2022-0186
    Cite
    @article{frau2023how,
  author    = {Frau, M. and Cabiddu, F. and Frigau, L. and Tomczyk, P. and Mola, F.},
  title     = {How emotions impact the interactive value formation process during problematic social media interactions},
  journal   = {Journal of Research in Interactive Marketing},
  year      = {2023},
  volume    = {17},
  number    = {5},
  pages     = {773--793},
  doi       = {10.1108/JRIM-06-2022-0186},
  issn      = {2040-7122},
}

  • Value Co-creation or Value Co-destruction? The Role of Negative Emotions in Consumer-Firm Interaction in the Social Media Platform

    M. Frau, L. Frigau, F. Cabiddu, F. Mola

    In The Palgrave Handbook of Interactive Marketing, pp. 987–1011. Palgrave Macmillan

    doi:10.1007/978-3-031-14961-0_42
    Cite
    @incollection{frau2023value,
  author    = {Frau, M. and Frigau, L. and Cabiddu, F. and Mola, F.},
  title     = {Value Co-creation or Value Co-destruction? The Role of Negative Emotions in Consumer-Firm Interaction in the Social Media Platform},
  booktitle = {The Palgrave Handbook of Interactive Marketing},
  editor    = {Wang, C. L.},
  publisher = {Palgrave Macmillan},
  address   = {Cham},
  year      = {2023},
  pages     = {987--1011},
  doi       = {10.1007/978-3-031-14961-0_42},
  isbn      = {978-3-031-14961-0},
}

  • Multivariate Regression Tree Topic Modeling

    M. Ortu, L. Frigau, G. Contu

    CLADAG 2023 Book of Abstracts and Short Papers, pp. 573–576

    Cite
    @inproceedings{ortu2023multivariate,
  author    = {Ortu, M. and Frigau, L. and Contu, G.},
  title     = {Multivariate Regression Tree Topic Modeling},
  booktitle = {CLADAG 2023 Book of Abstracts and Short Papers},
  year      = {2023},
  pages     = {573--576},
  isbn      = {978-88-919-3563-2},
}

2022

  • Optimizing the JSM Program

    L. Frigau, Q. Wu, D. Banks

    Journal of the American Statistical Association, 117(538): 617–626

    Seeded topic models and a scheduling algorithm reduce the overlap of topics between parallel sessions of the Joint Statistical Meetings.

    doi:10.1080/01621459.2021.1978466
    Cite
    @article{frigau2022optimizing,
  author    = {Frigau, L. and Wu, Q. and Banks, D.},
  title     = {Optimizing the JSM Program},
  journal   = {Journal of the American Statistical Association},
  year      = {2022},
  volume    = {117},
  number    = {538},
  pages     = {617--626},
  doi       = {10.1080/01621459.2021.1978466},
  issn      = {0162-1459},
}

  • Optimizing time slots in scientific meetings: a Latent Dirichlet allocation approach

    L. Frigau

    SIS 2022 Book of the Short Papers, pp. 110–120

    Cite
    @inproceedings{frigau2022optimizingb,
  author    = {Frigau, L.},
  title     = {Optimizing time slots in scientific meetings: a Latent Dirichlet allocation approach},
  booktitle = {SIS 2022 Book of the Short Papers},
  year      = {2022},
  pages     = {110--120},
  isbn      = {978-88-919-3231-0},
}

  • Explaining Student Satisfaction Assessments: A Natural Language Processing Approach

    M. Ortu, L. Frigau, G. Contu

    IES 2022 Book of Short Papers, pp. 313–318

    Cite
    @inproceedings{ortu2022explaining,
  author    = {Ortu, M. and Frigau, L. and Contu, G.},
  title     = {Explaining Student Satisfaction Assessments: A Natural Language Processing Approach},
  booktitle = {IES 2022 Book of Short Papers},
  year      = {2022},
  pages     = {313--318},
  isbn      = {978-88-94593-36-5},
}

2019

  • Customer Satisfaction from Booking

    M. Romano, L. Frigau, G. Contu, F. Mola, C. Conversano

    GARR_18 Conference - Data (R)evolution - Selected Papers, pp. 111–118. Associazione Consortium GARR

    doi:10.26314/GARR-Conf18-proceedings-21
    Cite
    @inproceedings{romano2019customer,
  author    = {Romano, M. and Frigau, L. and Contu, G. and Mola, F. and Conversano, C.},
  title     = {Customer Satisfaction from Booking},
  booktitle = {GARR\_18 Conference - Data (R)evolution - Selected Papers},
  publisher = {Associazione Consortium GARR},
  address   = {Rome},
  year      = {2019},
  pages     = {111--118},
  doi       = {10.26314/GARR-Conf18-proceedings-21},
  isbn      = {978-88-905077-8-6},
}

  • Value Co-Destruction: a Text-Mining-Based Mixed Method Study on Social Media Interactions

    M. Frau, L. Frigau, S. Serpi, F. Cabiddu

    Proceedings of the 20th International Conference Marketing Trends 2019 Venice. Paris-Venice Marketing Trends Association

    Cite
    @inproceedings{frau2019value,
  author    = {Frau, M. and Frigau, L. and Serpi, S. and Cabiddu, F.},
  title     = {Value Co-Destruction: a Text-Mining-Based Mixed Method Study on Social Media Interactions},
  booktitle = {Proceedings of the 20th International Conference Marketing Trends 2019 Venice},
  editor    = {Andreani, J.-C. and Collesei, U.},
  publisher = {Paris-Venice Marketing Trends Association},
  year      = {2019},
  isbn      = {978-2-490372-06-5},
}

More research stories