DLATK

Overview

DLATK (Differential Language Analysis ToolKit) is a Python toolkit for correlating language with outcome variables. Differential Language Analysis (DLA) is its signature method. DLA extracts language features, such as n-grams, from a corpus, then correlates the frequency of each feature with an outcome variable across participants.

We apply DLA to the same synthetic ChatbotLab corpus used in the other tutorials in this section: 19 person-to-AI conversations, each “person” tagged with a PHQ-9 depression score. We correlate single-word frequencies with PHQ-9.

Data Format

DLATK reads from a SQL database. It expects two tables: a message table (message_id, user_id, message) and an outcome table (user_id, outcome columns). DLATK defaults to MySQL, but it also supports SQLite, so no database server is required.

ChatbotLab exports conversation data in this shape already, as a SQLite database with a msgs table and an outcomes table. We use that export directly, at conversation_data/dlatk/dlatk.db.

The message table can hold many rows per participant, one per message. DLATK aggregates across whichever column is passed to -c (its alias is -g) when it extracts features and runs correlations. Messages do not need to be combined into one row per participant first.

Setup

pip install dlatk

Extracting Unigram Features

DLATK extracts n-gram features with --add_ngrams. We extract unigrams only, with -n 1. DLATK’s default is 1- to 3-grams, but with only 19 participants, most 2- and 3-grams occur too rarely to correlate with anything. Unigrams give more participants a non-zero count per word, which matters more than usual on a corpus this small:

dlatkInterface.py --db_engine sqlite -d conversation_data/dlatk/dlatk \
  -t msgs -c user_id \
  --add_ngrams -n 1

This creates the feature table feat$1gram$msgs$user_id inside dlatk.db.

Filtering Rare Features

Drop words used by too few participants, since a word only one person uses cannot be a reliable correlate of anything:

dlatkInterface.py --db_engine sqlite -d conversation_data/dlatk/dlatk \
  -t msgs -c user_id \
  -f 'feat$1gram$msgs$user_id' --feat_occ_filter --set_p_occ 0.05

This keeps 243 distinct unigrams, each used by at least 5% of participants.

Running Differential Language Analysis

Correlate each of the 243 unigrams with PHQ-9:

dlatkInterface.py --db_engine sqlite -d conversation_data/dlatk/dlatk \
  -t msgs -c user_id \
  -f 'feat$1gram$msgs$user_id$0_05' \
  --outcome_table outcomes --outcomes phq9 \
  --group_freq_thresh 50 \
  --correlate --csv --rmatrix --sort --no_correction \
  --output_name conversation_data/dlatk/dla_phq9

By default, DLATK corrects p-values for the number of features tested. With 243 features and 19 participants, that correction leaves nothing significant. --no_correction turns it off, so the p-values below are uncorrected. Treat them as exploratory, the same way we treat the z-scores and SDP scores in the other two tutorials.

Results

Words significantly correlated with higher PHQ-9, uncorrected p < .05:

word	r	p	freq
that	0.674	.0016	51
weeks	0.586	.0083	7
just	0.583	.0088	98
he	0.513	.0247	13
maybe	0.499	.0295	42
feels	0.497	.0303	17
it	0.464	.0453	140
make	0.461	.0467	6

Words significantly correlated with lower PHQ-9:

word	r	p	freq
one	-0.589	.0080	17
it’d	-0.568	.0111	4
phone’s	-0.550	.0147	3
gonna	-0.540	.0171	20
itself	-0.519	.0228	4
meaning	-0.510	.0256	5
actually	-0.503	.0281	37

just and feels correlate with higher PHQ-9 here, the same direction found by ConvoKit’s Fighting Words and the R text package’s Supervised Dimension Projection on this corpus. Three different methods point to the same words.

With only 19 participants and no p-value correction, this is a demonstration of the workflow, not a finding. Re-run this pipeline on a properly powered study export, and drop --no_correction.

Generalizing

Swap phq9 for any other column in outcomes.csv, such as age, or a REDCap or Qualtrics field passed through at Survey Integration. DLATK also extracts features beyond raw n-grams, including LIWC categories, LDA topics, and transformer embeddings, any of which can replace the unigram feature table in the correlation step.