Select the Proper One: Evaluating Subject Fashions for Enterprise Intelligence

are utilized in companies to categorise brand-related textual content datasets (similar to product and web site evaluations, surveys, and social media feedback) and to trace how buyer satisfaction metrics change over time.

There’s a myriad of latest subject fashions one can select from: the broadly used BERTopic by Maarten Grootendorst (2022), the latest FASTopic offered finally 12 months’s NeurIPS, (Xiaobao Wu et al.,2024), the Dynamic Subject Mannequin by Blei and Lafferty (2006), or a contemporary semi-supervised Seeded Poisson Factorization mannequin (Prostmaier et al., 2025).

For a enterprise use case, coaching subject fashions on buyer texts, we regularly get outcomes that aren’t similar and typically even conflicting. In enterprise, imperfections price 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 subject fashions seem in the marketplace, strategies for evaluating their high quality utilizing new metrics additionally evolve.

This sensible tutorial will concentrate on bigram subject fashions, which offer extra related data and establish 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 facet, bigram fashions are extra detailed; on the opposite, many analysis metrics weren’t initially designed for his or her analysis. To offer extra background on this space, we are going to discover intimately:

• Easy methods to consider the standard of bigram subject fashions
• Easy methods to put together an electronic mail classification pipeline in Python. 

Our instance use case will present how bigram subject fashions (BERTopic and FASTopic) assist prioritize electronic mail communication with clients on sure matters and scale back response occasions.

1. What are subject mannequin high quality indicators?

The analysis process ought to goal the perfect state:

The best subject mannequin ought to produce matters the place phrases or bigrams (two consecutive phrases) in every subject are extremely semantically associated and distinct for every subject.

In apply, because of this the phrases predicted for every subject are semantically related to human judgment, and there may be low duplication of phrases between matters. 

It’s normal to calculate a set of metrics for every educated mannequin to make a professional choice on which mannequin to position into manufacturing or use for a enterprise choice, evaluating the mannequin efficiency metrics.

• Coherence metrics consider how properly the phrases found by a subject mannequin make sense to people (have related semantics in every subject).
• Subject range measures how totally different the found matters are from each other. 

Bigram subject fashions work properly with these metrics:

• NPMI (Normalized Level-wise Mutual Info) makes use of chances 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 might be both inner (the coaching set) or exterior (e.g., an exterior electronic mail dataset). A big, exterior, and comparable corpus is a better option as a result of it might 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 combination mannequin rating is the typical of phrases throughout matters. 

• 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 have now the High 4 phrases for one subject: “apple”, “banana”, “juice”, “smoothie” predicted by a subject mannequin. Then SC appears in any respect mixtures 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. General 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 matters within the mannequin. PUV = 1 signifies that every subject within the mannequin incorporates distinctive bigrams. Values near 1 point out a well-shaped, high-quality mannequin with small phrase overlap between matters. [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 subject are semantically related). The nearer to 1 PUV is, the simpler the mannequin is to interpret and use, as a result of bigrams between matters don’t overlap. 

2. How can we prioritize electronic mail communication with subject fashions?

A big share of buyer communication, not solely in e-commerce companies, is now solved with chatbots and private shopper sections. But, it’s common to speak with clients by electronic mail. Many electronic mail suppliers supply builders broad flexibility in APIs to customise their electronic mail platform (e.g., MailChimp, SendGrid, Brevo). On this place, subject 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 subject classifier to categorize the incoming electronic mail content material. The end result is the categorized subject that the Buyer Care (CC) Division sees subsequent to every electronic mail. The primary 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 straight 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 matters and consider the standard of all mannequin specs. Learn my earlier TDS tutorial on subject modeling with these cutting-edge subject fashions.

As a coaching set, we use a synthetically generated Buyer Care E mail dataset out there on Kaggle with a GPL-3 license. The prefiltered information covers 692 incoming emails and appears like this:

3.1. Information preprocessing

Cleansing textual content in the fitting order is crucial for subject fashions to work in apply as a result of it minimizes the bias of every cleansing operation. 

Numbers are sometimes eliminated first, adopted by emojis, until we don’t want them for particular conditions, similar 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 information 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, subject fashions would mannequin them as separate components. 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 information on enter; some cleansing (stopwords) might be executed in the course of the coaching. The only and only approach is to make use of the Washer, a no-code app for textual content information cleansing that gives a no-code approach of information preprocessing for textual content mining initiatives.
• BERTopic: the documentation recommends that “removing cease phrases as a preprocessing step shouldn’t be suggested because the transformer-based embedding fashions that we use want the complete context to create correct embeddings”. Because of this, cleansing operations needs to be included within the mannequin coaching.

3.2. Mannequin compilation and coaching

You possibly can test 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 matters in buyer electronic mail information. A easy inspection of the subject distribution exhibits that incoming emails to FASTopic are fairly properly distributed throughout matters. BERTopic classifies emails erratically, maintaining outliers (uncategorized tokens) in T-1 and a big share of incoming emails in T0.

Listed below are the expected bigrams for each fashions with subject labels:

As a result of the e-mail corpus is an artificial LLM-generated dataset, the naive labelling of the matters for each fashions exhibits matters which are:

• Comparable: Time Delays, Latency Points, Consumer Permissions, Deployment Points, Compilation Errors,
• Differing: Unclassified (BERTopic classifies outliers into T-1), Enchancment Options, Authorization Errors, Efficiency Complaints (FASTopic), Cloud Administration, Asynchronous Requests, Basic Requests (BERTopic)

For enterprise functions, matters needs to be labelled by the corporate’s insiders who know the shopper base and the enterprise priorities.

4. Mannequin analysis

If three out of eight categorized matters are labeled in a different way, 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 Buyer IT Help Ticket Dataset from Kaggle, distributed with Attribution 4.0 Worldwide license, gives comparable information to our coaching set. The information is filtered to 11923 English electronic mail our bodies. 

1. Calculate an NPMI for every bigram within the reference corpus with this code.
2. 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 matters 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:

1. Create a document-term matrix (DTM) with a depend of what number of occasions every bigram seems in every doc.
2. Calculate subject SC scores by trying to find bigram co-occurrences within the DTM and the bigrams predicted by subject fashions.
3. Common subject SC to a mannequin SC rating.

4.3. PUV

Subject range PUV metric checks the duplicates of bigrams between matters in a mannequin.

1. Be part of bigrams into tokens by changing areas with underscores within the FASTopic and BERTopic tables of predicted bigrams.

2. Calculate subject range 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 numerous than FASTopic. The variations will not be very giant, 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 subject. Growing this parameter causes the mannequin to be unable to coach, most likely due to the small information measurement. 

Because of this, FASTopic is a better option for subject modelling in electronic mail classification with small coaching datasets.

The final step is to deploy the mannequin with subject labels within the electronic mail platform to categorise incoming emails:

Abstract

Coherence and variety metrics examine fashions with related coaching setups, the identical dataset, and cleansing technique. We can not examine their absolute values with the outcomes of various coaching classes. 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 apply.

How does buyer care profit from the subject modelling train? After the subject mannequin is put into manufacturing, the pipeline sends a categorized subject for every electronic mail to the e-mail platform that Buyer Care makes use of for speaking with clients. With a restricted employees, it’s now doable to prioritize and reply sooner to probably the most delicate enterprise requests (similar to “time delays” and “latency points”), and alter priorities dynamically. 

Information and full codes for this tutorial are right here.

Petr Korab is a Python Engineer and Founding father of Textual content Mining Tales 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 subject 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. Subject Modeling in embedding areas. Transactions of the Affiliation for Computational Linguistics, 8:439-453.

[3] Grootendorst, M. 2022. Bertopic: Neural Subject Modeling With A Class-Primarily based TF-IDF Process. Laptop Science.

[4] Korab, P. Subject Modelling in Enterprise Intelligence: FASTopic and BERTopic in Code. In direction of Information Science. 22.1.2025. Accessible from: hyperlink.

[5] Korab, P. Subject Modelling with BERTtopic in Python. In direction of Information Science. 4.1.2024. Accessible from: hyperlink.

[6] Wu, X, Nguyen, T., Ce Zhang, D., Yang Wang, W., Luu, A. T. 2024. FASTopic: A Quick, Adaptive, Secure, and Transferable Subject Modeling Paradigm. arXiv preprint: 2405.17978.

[7] Mimno, D., Wallach, H., M., Talley, E., Leenders, M, McCallum. A. 2011. Optimizing Semantic Coherence in Subject Fashions. 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 area data to suit subject fashions. arXiv preprint: 2405.17978.