How do I set up topic modeling software for extracting themes from support tickets?

Here’s a practical way to set up topic modeling for support tickets, from raw ticket text to usable themes.

1) Clarify the goal

Decide what you want the model to produce:

• Broad themes: “login issues,” “billing problems,” “feature requests”
• Operational categories: “password reset,” “payment failure,” “bug report”
• Trend monitoring: how themes change over time
• Routing/triage support: help agents classify incoming tickets

For support tickets, the best results usually come from topic modeling plus some manual labeling, not fully automatic topics alone.

---

2) Choose the right approach

Classic topic modeling often uses:

• LDA (Latent Dirichlet Allocation)
  Good for large corpora, but can be weak on short texts like tickets.

Better options for support tickets:

• BERTopic
  Uses embeddings + clustering; often works much better for short, messy text.
• NMF (Non-negative Matrix Factorization)
  Simple and often effective with TF-IDF.
• Guided / semi-supervised topic modeling
  If you already know some themes, you can steer the model.

If your tickets are short and varied, I’d usually start with BERTopic.

---

3) Collect and clean the data

Export ticket data with at least:

• ticket ID
• subject/title
• body/description
• tags or category if available
• timestamp
• resolution / status

Cleaning steps

• Remove duplicates
• Strip signatures, disclaimers, and canned text
• Combine subject + body
• Normalize casing
• Remove obvious noise:
  • emails
  • URLs
  • tracking IDs
  • long signatures
• Consider anonymizing personal data

Example text after cleanup:

“Unable to reset password after clicking email link. Error says token expired.”

---

4) Preprocess carefully

For topic modeling, you typically want:

• Tokenization
• Lowercasing
• Stopword removal
• Lemmatization or stemming
• Optional bigrams/trigrams

But don’t over-clean support tickets. Important phrases like:

• “credit card”
• “log in”
• “two factor”
• “error code”

should be preserved as multi-word terms if possible.

Useful extra steps:

• Add domain stopwords like:
  • “please”
  • “thank”
  • “help”
  • “ticket”
  • “issue”
  • “customer”
• Preserve product names and feature names

---

5) Pick software

Common options:

Python

Best flexibility.

Libraries:

• BERTopic
• scikit-learn
• gensim
• spaCy
• sentence-transformers

Typical stack for support tickets:

• pandas for data
• spaCy for preprocessing
• sentence-transformers for embeddings
• BERTopic for modeling

No-code / low-code tools

If you want something easier:

• KNIME
• RapidMiner
• Orange
• BI platforms with text analytics add-ons

These are faster to start with, but less flexible.

---

6) Example setup with BERTopic in Python

A minimal workflow:

import pandas as pd
from bertopic import BERTopic

# Load data
df = pd.read_csv("tickets.csv")
docs = (df["subject"].fillna("") + " " + df["body"].fillna("")).tolist()

# Build model
topic_model = BERTopic(language="english", min_topic_size=20)
topics, probs = topic_model.fit_transform(docs)

# View topics
topic_model.get_topic_info()

Useful settings to tune

• min_topic_size: controls how small a topic can be
• nr_topics: reduce number of topics if too many are created
• calculate_probabilities=True: for confidence scores
• custom vectorizer_model: for stopwords and n-grams

Example with custom n-grams:

from sklearn.feature_extraction.text import CountVectorizer

vectorizer_model = CountVectorizer(
    stop_words="english",
    ngram_range=(1, 2),
    min_df=5
)

topic_model = BERTopic(vectorizer_model=vectorizer_model, min_topic_size=20)

---

7) Evaluate the topics

Topic modeling is only useful if humans can interpret the results.

Check:

• Do the top words make sense?
• Are topics distinct or too overlapping?
• Are there too many “miscellaneous” tickets?
• Do themes align with known support categories?

Practical evaluation methods

• Manually review top 20 tickets per topic
• Compare with existing support labels
• Measure coherence if available
• Ask support agents whether topics are meaningful

---

8) Refine the model

Common refinements:

• Add domain stopwords
• Merge similar topics
• Remove very rare terms
• Use better embeddings
• Separate by product line or ticket type
• Filter by language if multilingual

BERTopic supports topic reduction and merging, which is useful if the output is too fragmented.

---

9) Operationalize it

Once the model is good enough:

• Run it on new tickets daily or weekly
• Assign topic labels automatically
• Track topic volume over time
• Build dashboards for:
  • top emerging issues
  • spikes in complaints
  • recurring bugs
• Use it to support escalation and root-cause analysis

---

10) Recommended starting architecture

A solid, practical setup:

1. Export tickets from Zendesk/Jira/Freshdesk/etc.
2. Clean and normalize text
3. Combine subject + body
4. Use BERTopic with custom stopwords
5. Review and manually name topics
6. Save topic labels to your ticket system or warehouse
7. Monitor topic trends in a dashboard

---

Quick recommendation

If you want the simplest strong solution for support tickets:

• Use Python
• Start with BERTopic
• Add domain-specific stopwords
• Review topics manually
• Use the topics for reporting and trend detection, not as final truth

If you want, I can also give you:

1. a step-by-step Python setup guide,
2. a sample notebook, or
3. advice tailored to Zendesk/Jira/Freshdesk tickets.