are utilized in companies to categorise brand-related textual content datasets (corresponding to product and web site opinions, surveys, and social media feedback) and to trace how buyer satisfaction metrics change over time.
There’s a myriad of current matter fashions one can select from: the extensively used BERTopic by Maarten Grootendorst (2022), the current FASTopic introduced eventually yr’s NeurIPS, (Xiaobao Wu et al.,2024), the Dynamic Topic Model by Blei and Lafferty (2006), or a contemporary semi-supervised Seeded Poisson Factorization mannequin (Prostmaier et al., 2025).
For a enterprise use case, coaching matter fashions on buyer texts, we frequently get outcomes that aren’t equivalent and typically even conflicting. In enterprise, imperfections value cash, so the engineers ought to place into manufacturing the mannequin that gives the most effective answer and solves the issue most successfully. On the similar tempo that new matter fashions seem in the marketplace, strategies for evaluating their high quality utilizing new metrics additionally evolve.
This sensible tutorial will concentrate on bigram matter fashions, which offer extra related info and determine higher key qualities and issues for enterprise choices than single-word fashions (“supply” vs. “poor supply”, “abdomen” vs. “delicate abdomen”, and so forth.). On one aspect, bigram fashions are extra detailed; on the opposite, many analysis metrics weren’t initially designed for his or her analysis. To supply extra background on this space, we’ll discover intimately:
- Easy methods to consider the standard of bigram matter fashions
- Easy methods to put together an e-mail classification pipeline in Python.
Our instance use case will present how bigram matter fashions (BERTopic and FASTopic) assist prioritize e-mail communication with prospects on sure subjects and scale back response instances.
1. What are matter mannequin high quality indicators?
The analysis activity ought to goal the best state:
The best matter mannequin ought to produce subjects the place phrases or bigrams (two consecutive phrases) in every matter are extremely semantically associated and distinct for every matter.
In follow, because of this the phrases predicted for every matter are semantically similar to human judgment, and there may be low duplication of phrases between subjects.
It’s normal to calculate a set of metrics for every educated mannequin to make a certified resolution on which mannequin to put into manufacturing or use for a enterprise resolution, evaluating the mannequin efficiency metrics.
- Coherence metrics consider how properly the phrases found by a subject mannequin make sense to people (have similar semantics in every matter).
- Subject variety measures how totally different the found subjects are from each other.
Bigram matter fashions work properly with these metrics:
- NPMI (Normalized Level-wise Mutual Data) makes use of possibilities estimated in a reference corpus to calculate a [-1:1] rating for every phrase (or bigram) predicted by the mannequin. Learn [1] for extra particulars.
The reference corpus may be both inner (the coaching set) or exterior (e.g., an exterior e-mail dataset). A big, exterior, and comparable corpus is a more sensible choice as a result of it will probably assist scale back bias in coaching units. As a result of this metric works with phrase frequencies, the coaching set and the reference corpus needs to be preprocessed the identical approach (i.e., if we take away numbers and stopwords within the coaching set, we must also do it within the reference corpus). The mixture mannequin rating is the typical of phrases throughout subjects.
- SC (Semantic Coherence) doesn’t want a reference corpus. It makes use of the identical dataset as was used to coach the subject mannequin. Learn extra in [2].
Let’s say we now have the Prime 4 phrases for one matter: “apple”, “banana”, “juice”, “smoothie” predicted by a subject mannequin. Then SC appears in any respect combos of phrases within the coaching set going from left to proper, beginning with the primary phrase {apple, banana}, {apple, juice}, {apple, smoothie} then the second phrase {banana, juice}, {banana, smoothie}, then final phrase {juice, smoothie} and it counts the variety of paperwork that include each phrases, divided by the frequency of paperwork that include the primary phrase. Total SC rating for a mannequin is the imply of all topic-level scores.
PUV (Share of Distinctive Phrases) calculates the share of distinctive phrases throughout subjects within the mannequin. PUV = 1 signifies that every matter within the mannequin incorporates distinctive bigrams. Values near 1 point out a well-shaped, high-quality mannequin with small phrase overlap between subjects. [3].
The nearer to 0 the SC and NIMP scores are, the extra coherent the mannequin is (bigrams predicted by the subject mannequin for every matter are semantically related). The nearer to 1 PUV is, the better the mannequin is to interpret and use, as a result of bigrams between subjects don’t overlap.
2. How can we prioritize e-mail communication with matter fashions?
A big share of buyer communication, not solely in e-commerce companies, is now solved with chatbots and private consumer sections. But, it’s common to speak with prospects by e-mail. Many e-mail suppliers provide builders broad flexibility in APIs to customise their e-mail platform (e.g., MailChimp, SendGrid, Brevo). On this place, matter fashions make mailing extra versatile and efficient.
On this use case, the pipeline takes the enter from the incoming emails and makes use of the educated matter classifier to categorize the incoming e-mail content material. The end result is the categorized matter that the Buyer Care (CC) Division sees subsequent to every e-mail. The principle goal is to permit the CC employees to prioritize the classes of emails and scale back the response time to probably the most delicate requests (that instantly have an effect on margin-related KPIs or OKRs).
3. Information and mannequin set-ups
We are going to prepare FASTopic and Bertopic to categorise emails into 8 and 10 subjects and consider the standard of all mannequin specs. Learn my earlier TDS tutorial on matter modeling with these cutting-edge matter fashions.
As a coaching set, we use a synthetically generated Customer Care Email dataset out there on Kaggle with a GPL-3 license. The prefiltered knowledge covers 692 incoming emails and appears like this:
3.1. Information preprocessing
Cleansing textual content in the fitting order is important for matter fashions to work in follow as a result of it minimizes the bias of every cleansing operation.
Numbers are usually eliminated first, adopted by emojis, until we don’t want them for particular conditions, corresponding to extracting sentiment. Stopwords for a number of languages are eliminated afterward, adopted by punctuation in order that stopwords don’t break up into two tokens (“we’ve” -> “we” + ‘ve”). Extra tokens (firm and folks’s names, and so forth.) are eliminated within the subsequent step within the clear knowledge earlier than lemmatization, which unifies tokens with the identical semantics.
“Supply” and “deliveries”, “field” and “Packing containers”, or “Worth” and “costs” share the identical phrase root, however with out lemmatization, matter fashions would mannequin them as separate elements. That’s why buyer emails needs to be lemmatized within the final step of preprocessing.
Textual content preprocessing is model-specific:
- FASTopic works with clear knowledge on enter; some cleansing (stopwords) may be accomplished throughout the coaching. The only and only approach is to make use of the Washer, a no-code app for text data cleaning that provides a no-code approach of information preprocessing for textual content mining tasks.
- BERTopic: the documentation recommends that “removing cease phrases as a preprocessing step will not be suggested because the transformer-based embedding fashions that we use want the complete context to create correct embeddings”. For that reason, cleansing operations needs to be included within the mannequin coaching.
3.2. Mannequin compilation and coaching
You possibly can verify the complete codes for FASTopic and BERTopic’s coaching with bigram preprocessing and cleansing in this repo. My earlier TDS tutorials (4) and (5) clarify all steps intimately.
We prepare each fashions to categorise 8 subjects in buyer e-mail knowledge. A easy inspection of the subject distribution exhibits that incoming emails to FASTopic are fairly properly distributed throughout subjects. BERTopic classifies emails inconsistently, retaining outliers (uncategorized tokens) in T-1 and a big share of incoming emails in T0.
Listed here are the expected bigrams for each fashions with matter labels:
As a result of the e-mail corpus is an artificial LLM-generated dataset, the naive labelling of the subjects for each fashions exhibits subjects which might be:
- Comparable: Time Delays, Latency Points, Consumer Permissions, Deployment Points, Compilation Errors,
- Differing: Unclassified (BERTopic classifies outliers into T-1), Enchancment Solutions, Authorization Errors, Efficiency Complaints (FASTopic), Cloud Administration, Asynchronous Requests, Basic Requests (BERTopic)
For enterprise functions, subjects needs to be labelled by the corporate’s insiders who know the client base and the enterprise priorities.
4. Mannequin analysis
If three out of eight categorized subjects are labeled otherwise, then which mannequin needs to be deployed? Let’s now consider the coherence and variety for the educated BERTopic and FASTopic T-8 fashions.
4.1. NPMI
We’d like a reference corpus to calculate an NPMI for every mannequin. The Customer IT Support Ticket Dataset from Kaggle, distributed with Attribution 4.0 International license, supplies comparable knowledge to our coaching set. The information is filtered to 11923 English e-mail our bodies.
- Calculate an NPMI for every bigram within the reference corpus with this code.
- Merge bigrams predicted by FASTopic and BERTopic with their NPMI scores from the reference corpus. The less NaNs are within the desk, the extra correct the metric is.
3. Common NPMIs inside and throughout subjects to get a single rating for every mannequin.
4.2. SC
With SC, we be taught the context and semantic similarity of bigrams predicted by a subject mannequin by calculating their place within the corpus in relation to different tokens. To take action, we:
- Create a document-term matrix (DTM) with a depend of what number of instances every bigram seems in every doc.
- Calculate matter SC scores by looking for bigram co-occurrences within the DTM and the bigrams predicted by matter fashions.
- Common matter SC to a mannequin SC rating.
4.3. PUV
Subject variety PUV metric checks the duplicates of bigrams between subjects in a mannequin.
- Be a part of bigrams into tokens by changing areas with underscores within the FASTopic and BERTopic tables of predicted bigrams.
2. Calculate matter variety as depend of distinct tokens/ depend of tokens within the tables for each fashions.
4.4. Mannequin comparability
Let’s now summarize the coherence and variety analysis in Picture 9. BERTopic fashions are extra coherent however much less various than FASTopic. The variations should not very massive, however BERTopic suffers from uneven distribution of incoming emails into the pipeline (see charts in Picture 5). Round 32% of categorized emails fall into T0, and 15% into T-1, which covers the unclassified outliers. The fashions are educated with a min. of 20 tokens per matter. Rising this parameter causes the mannequin to be unable to coach, most likely due to the small knowledge dimension.
For that reason, FASTopic is a more sensible choice for matter modelling in e-mail classification with small coaching datasets.
The final step is to deploy the mannequin with matter labels within the e-mail platform to categorise incoming emails:
Abstract
Coherence and variety metrics evaluate fashions with related coaching setups, the identical dataset, and cleansing technique. We can not evaluate their absolute values with the outcomes of various coaching periods. However they assist us resolve on the most effective mannequin for our particular use case. They provide a relative comparability of assorted mannequin specs and assist resolve which mannequin needs to be deployed within the pipeline. Subject fashions analysis ought to at all times be the final step earlier than mannequin deployment in enterprise follow.
How does buyer care profit from the subject modelling train? After the subject mannequin is put into manufacturing, the pipeline sends a categorized matter for every e-mail to the e-mail platform that Buyer Care makes use of for speaking with prospects. With a restricted employees, it’s now attainable to prioritize and reply quicker to probably the most delicate enterprise requests (corresponding to “time delays” and “latency points”), and alter priorities dynamically.
Information and full codes for this tutorial are here.
Petr Korab is a Python Engineer and Founding father of Text Mining Stories with over eight years of expertise in Enterprise Intelligence and NLP.
Acknowledgments: I thank Tomáš Horský (Lentiamo, Prague), Martin Feldkircher, and Viktoriya Teliha (Vienna Faculty of Worldwide Research) for helpful feedback and ideas.
References
[1] Blei, D. M., Lafferty, J. D. 2006. Dynamic matter fashions. In Proceedings of the twenty third worldwide convention on Machine studying (pp. 113–120).
[2] Dieng A.B., Ruiz F. J. R., and Blei D. M. 2020. Topic Modeling in embedding areas. Transactions of the Association for Computational Linguistics, 8:439-453.
[3] Grootendorst, M. 2022. Bertopic: Neural Subject Modeling With A Class-Primarily based TF-IDF Process. Computer Science.
[4] Korab, P. Subject Modelling in Enterprise Intelligence: FASTopic and BERTopic in Code. In direction of Information Science. 22.1.2025. Accessible from: link.
[5] Korab, P. Subject Modelling with BERTtopic in Python. In direction of Information Science. 4.1.2024. Accessible from: link.
[6] Wu, X, Nguyen, T., Ce Zhang, D., Yang Wang, W., Luu, A. T. 2024. FASTopic: A Fast, Adaptive, Stable, and Transferable Topic Modeling Paradigm. arXiv preprint: 2405.17978.
[7] Mimno, D., Wallach, H., M., Talley, E., Leenders, M, McCallum. A. 2011. Optimizing Semantic Coherence in Topic Models. Proceedings of the 2011 Convention on Empirical Strategies in Pure Language Processing.
[8] Prostmaier, B., Vávra, J., Grün, B., Hofmarcher., P. 2025. Seeded Poisson Factorization: Leveraging domain knowledge to fit topic models. arXiv preprint: 2405.17978.
