LLM Classification

Classify concordance lines with custom variables

Overview

LLM classification lets you assign linguistic labels to a random sample of concordance lines. You define variables (enum or open) and the model returns one label per variable.

Presets

Use preset variables for common tasks:

  • Subject animacy (animate vs inanimate)
  • Transitivity (transitive vs intransitive vs ditransitive)
  • Topic (open, coarse-grained domain)
  • Illocution (Searle taxonomy)

Custom variables

Create your own variables:

  • Enum: fixed set of options
  • Open: free-text label (1-3 words recommended)

Add a short description and optional examples to guide the model.

How to run

  1. Run a CQL query in the custom app.
  2. Open LLM classification (bottom of the page).
  3. Add presets and/or custom variables.
  4. Choose a sample size and click Classify sample.

Output

  • A table with Left / KWIC / Right + your variable columns.
  • CSV export with the same columns.

Notes & tips

  • Keep variable descriptions concise and concrete.
  • Use examples to indicate the intended granularity.
  • The LLM uses the server OpenAI key (no client key required).

Linguistic Use Cases

Beyond the preset variables, LLM classification enables diverse linguistic analyses:

Word Sense Disambiguation

Classify polysemous words by their contextual meaning:

  • Query: [lemma="strana"] in skRed (Slovak)
  • Variable: Enum with options “spatial side”, “political party”
  • Result: Automatic sense labeling that complements semantic clustering

Grammatical Aspect (Slavic Languages)

Distinguish perfective vs. imperfective aspect:

  • Query: [lemma="robiť" | lemma="urobiť"] in skRed
  • Variable: Enum with “imperfective (process)”, “perfective (result)”
  • Application: Aspect studies, cross-linguistic comparison

Sentiment Analysis

Analyze attitudes toward entities in discourse:

  • Query: [word="Fico" | word="Ficovi" | word="Fica"] in skRed
  • Variable: Enum with “positive”, “neutral”, “negative”
  • Application: Political discourse analysis, public opinion research

Morphological Function

Classify productive morphological patterns:

  • Query: [word=".*[íč]ko"] in skRed (Slovak diminutives)
  • Variable: Enum with “actual diminution”, “affective/endearment”, “lexicalized”
  • Application: Morphology research, pragmatics

Causative Alternation

Distinguish argument structure patterns:

  • Query: [lemma="break" & tag="VV.*"] in COCA
  • Variable: Enum with “causative (X breaks Y)”, “inchoative (Y breaks)”
  • Application: Valency studies, construction grammar

Example: Classifying subject animacy in UniPlans

This walkthrough demonstrates LLM classification using the UniPlans corpus to analyze the grammatical subjects of “partnership” constructions.

Step 1: Run a query

First, run a CQL query to get concordance results:

  1. Select UniPlans (383K tokens) from the corpus dropdown.
  2. Query for [lemma=="partnership"].
  3. Click Run builder query.

Step 2: Configure variables

Scroll to the LLM classification panel. You’ll see:

  • Sample size: how many concordances to classify (5-100)
  • Estimated cost: calculated based on variables and sample size
  • Preset variable buttons: Subject animacy, Transitivity, Topic, Illocution

LLM classification panel with preset variable buttons

Open the sample size dropdown to pick how many lines to classify:

Sample size options in the classification panel

Click + Subject animacy to add the preset. The variable card appears with:

  • Title: Subject animacy
  • Description: Classification criteria with examples
  • Type: Fixed set of options (enum)
  • Options: animate, inanimate

Variable configuration showing description and enum options

The description provides guidance to the LLM:

Classify the grammatical subject’s animacy. animate = humans, animals, or organizations acting as agents; inanimate = objects, abstract concepts, events, or natural phenomena.

Examples help calibrate the model:

  • “The child laughed loudly.” -> animate
  • “The committee approved the proposal.” -> animate
  • “The storm destroyed the docks.” -> inanimate

Step 3: Run classification

  1. Select sample size (e.g., 5 lines for a quick test).
  2. Review the estimated cost.
  3. Click Classify sample.

The model classifies each concordance line and returns results.

Step 4: Review results

Results appear in a table with your variable columns:

Classification progress indicator while the job runs

Classification results showing concordances with Subject animacy annotations

Each row shows:

  • Left context: words before the keyword
  • KWIC: the keyword in context (partnerships)
  • Right context: words after the keyword
  • Subject animacy: the LLM’s classification

In this example, most “partnership” usages have animate subjects (organizations, institutions, “we”) because the corpus discusses what universities do with partnerships.

Step 5: Export results

Click Export CSV to download results for further analysis in Excel, R, or Python.

Variable design tips

Good variable definitions lead to better classifications. Follow these principles:

Be specific and concrete

Good: “Classify whether the subject is a human, animal, organization (animate) or an object, concept, event (inanimate).”

Vague: “Is the subject alive?”

Provide representative examples

Include 3-5 examples that show edge cases:

Examples:
- "The university invested heavily" → animate (organization as agent)
- "Our partnerships enable growth" → animate (implied "we")
- "The program achieved results" → inanimate (program as abstract)

Keep descriptions concise

The LLM has limited context. Descriptions under 100 words work best. Focus on:

  1. What to classify
  2. How to distinguish categories
  3. Key examples

Choose appropriate variable types

  • Enum (fixed options): Use when categories are mutually exclusive and well-defined. Better for consistency across samples.
  • Open (free text): Use for exploratory analysis or when categories are hard to predefine. Allows 1-6 word labels.

Test with small samples first

Run 5-10 classifications to verify the model understands your variable definition before scaling up.