TAALCR — Toolkit for Aggregate Analysis of Language in Conversation, for Research

Open-source Python infrastructure for reproducible conversational discourse analysis.

---

⚠️ Development Notice

Active development of the semi-automated discourse analysis infrastructure now continues under the DIAAD project.

This repository remains available for historical continuity and will later host modules for clinical language elicitation evaluation in aphasiology.

---

Description

TAALCR is a research toolkit for batched conversational discourse analysis.
It currently provides workflows for:

• Digital Conversation Turn (DCT) analysis
• POWERS coding for pragmatic discourse evaluation

TAALCR complements the monologic discourse analysis system
RASCAL and imports components from it when appropriate.

A third functionality — quantitative characterization of clinical language elicitation — is currently in development.

---

Installation

TAALCR is currently developed and tested with Python 3.12.

A dedicated virtual environment using Anaconda is recommended:

1. Create and activate your environment:

conda create --name taalcr python=3.12
conda activate taalcr

2. Install TAALCR:

# Install from PyPI
pip install taalcr

# or install the latest development version
pip install git+https://github.com/nmccloskey/taalcr.git@main

3. Install the spaCy transformer model

(required for automated POWERS coding)

python -m spacy download en_core_web_trf

---

Overview (more details below)

• Digital Conversation Turns Analysis
  • Tracking turn-taking in dialogs can reveal meaningful linguistic and psychosocial patterns (Tuomenoksa, et al., 2020).
  • Recording turns with a sequence of digits enables analysis of both tallies and transition probabilities (see below).
• POWERS Coding
  • Profile of Word Errors and Retrieval in Speech (POWERS) is an aphasiological coding system for analyzing dialogic speech (Herbet, et al., 2013).
  • TAALCR POWERS pipeline:
    • generates coder workbooks, automating most fields
    • summarizes coding and reports ICC2 values between coders
    • evaluates and optionally reselects reliability coding
  • Automation validation (CLI only)
    • select (stratified) random subset for manual coding
    • evaluate reliability between automatic & manual codes

---

Web App

You can use TAALCR in your browser — no installation required:

👉 Launch the TAALCR Web App

---

Setup

To prepare for running TAALCR, complete the following steps:

1. Create your working directory:

Example structure:

your_project/
├── config.yaml           # Configuration file (see below)
└── taalcr_data/
    └── input/            # Place your .cha or .xlsx files here
                          # (TAALCR will make an output directory)

2. Provide a config.yaml file

This file specifies the directories, coders, settings, and tier structure.

You can download the example config file from the repo or create your own like this:

input_dir: taalcr_data/input
output_dir: taalcr_data/output
reliability_fraction: 0.2
automate_POWERS: true
just_c2_POWERS: false
exclude_participants:
coders:
- '1'
- '2'
- '3'
tiers:
  time:
    values:
    - PreTx
    - PostTx
  client_id:
    values: \d+
  setting:
    values:
    - LargeGroup
    - SmallGroup

Explanation:

• General

  • reliability_fraction – proportion of data to subset for reliability (default 20%).
  • coders – alphanumeric coder identifiers (3 required for function powers make).
  • exclude_participants – speakers appearing in .cha files to exclude from POWERS coding files.
  • automate_POWERS – toggle automated preparation of POWERS coding spreadsheets (coder 1 fields).
  • just_c2_POWERS – whether to use only coder 2 columns in analysis outputs.

• Tiers

  • Define metadata fields extracted from filenames (time, client_id, setting).
  • Each tier has attributes:
    • values – acceptable set of identifiers or regex patterns.
    • partition – (True/False) creates separate coding and reliability files for that tier.

  See RASCAL for more information about the tier system for organizing data based on .cha file names.

---

Quickstart — Command Line

TAALCR exposes a concise CLI with subcommands:

# Analyze digital conversation turns
taalcr turns

# POWERS workflow
taalcr powers make       # prepare POWERS coding files
taalcr powers analyze    # analyze completed POWERS coding
taalcr powers evaluate   # evaluate completed POWERS reliability coding
taalcr powers reselect   # randomly reselect reliability subset

# Automation validation
taalcr powers select     # randomly select subset for validating automation
taalcr powers validate   # compute reliability metrics on automated vs manual codes 

---

Digital Conversation Turns (DCT) Protocol

TAALCR includes a lightweight system for analyzing digital conversational turns in group treatment sessions for people with aphasia.
Instead of simple tallies, the DCT protocol records the sequence of turns compactly, enabling analysis of turn-taking dynamics and engagement, with optional markers for capturing turn qualities (e.g., length/substantiveness).

---

Coding Procedure

1. Speaker Assignment

• 0 = Clinician(s) (all individuals not receiving treatment collapsed under this code)
• 1 = Participant 1
• 2 = Participant 2
• Continue incrementing (3, 4, …) as needed.

2. Turn Entry with Markers

For each conversational turn, enter the assigned digit for the speaker (e.g., 0, 1, 2).

Marking system:

• Digits are followed by one dot . (mark1), two dots .. (mark2) or no dots
• Example usage:
  • Add . if the turn is substantial (contains an independent clause).
  • Add .. if the turn is monologic (contains at least two independent clauses)
  • Add no dots otherwise, or the turn is minimal (brief/no full idea)

3. Input Coding Table Format

• Turns are entered sequentially as a continuous string of digits and dots.
• Bins are recommended for some temporal granularity (e.g., six 10-minute bins for a 1-hour conversation/treatment session).
• Case-insensitive file name regex r'.*(Convo|Conversation)_?Turns.*\.xlsx$' looks for files like *TU_ConvoTurns.xlsx or *converstation_turns_2025.xlsx

Example: Digital Conversation Turns Coding Input

site	session	group	coder	bin	turns
TU	12	Dyad1	NM	1	212012.02121210.10101.210.12.021212121210.210.2.1.010121.010.110.2102.12.
TU	12	Dyad1	NM	2	0202.121212101.011101.2.12.120201.212101020202.10.21212.02.12010212.
TU	12	Dyad1	NM	3	12..121.212.1212.0202.12120.201.210101..2012121.2121.2..1212.12.020.2.0
TU	12	Dyad1	NM	4	010202.02121021020212101.01012101210010102.1210101010101010101010121020.1.
TU	12	Dyad1	NM	5	0.121210.1010102120.102.02120212.0.2.020212121202121212.120.21010101212121
TU	12	Dyad1	NM	6	2120210101212121212.10121202.12.02.1212010202.02.02.0202.020201202020.22.02012102002.012102
TU	4	LgGroup	NM	1	4.24.242424.0640.4.206.434343430606.060436.3706.0406.76760.602.502.326207.07.67.06767.3737.17.0701270606.06.54321007
TU	4	LgGroup	NM	2	763670.50505620507102..02404676.70101...010.707057574767.6..76717.01.7010141.4..1014.3401.671..61016161.721.77414.0
TU	4	LgGroup	NM	3	2.0.2.0.3.13.23.01313535737037.0.7.137314.
TU	4	LgGroup	NM	4	4.0.5.35.05.0.5..7575404.53436..40575754..24242..575.4375.45705.20.6.
TU	4	LgGroup	NM	5	06.007070767676050.21627.17.106063434607571270101.61.01016.161.2.0.1.01
TU	4	LgGroup	NM	6	0.607.2707.07.06..06.06.4603403212607201202..2702760276..020.1212606016..70.701702.1.70731313510.

---

Output

The taalcr turns command analyzes coded conversation turn files and produces an Excel workbook with multiple sheets, capturing turn-taking behavior at bin, speaker, session, and group levels, also including transition matrices for a detailed view of conversational dynamics.

Excel Sheet	Level of Analysis	Data Included
Speaker_Level_Turns	Speaker	Total turns, dot-mark counts (mark1/mark2), proportions
Group_Level_Summary	Group	Group totals, num participants, num sessions, marker proportions
Session_Level_Summary	Session × Group	Totals, entropy, clinician–participant ratio, marker proportions
Participation_Level_Turns	Speaker × Session	Individual totals, session proportion, marker rates, bin variability stats
Bin_Level_Turns	Speaker × Bin	Proportion of bin turns, marker proportions within bins
Speaker_Matrix_ *	Group	Conditional probabilities of turn transitions (matrix)
Speaker_Level_Ratios	Group	Participant→Participant, Participant→Clinician, Clinician→Participant ratios
Summary_Statistics	Aggregated	Mean, std, min, max, CV for all numeric metrics

---

Analytic Opportunities

• Turn counts & proportions per participant
• Substantial vs. monologic vs. minimal turn ratios
• Transitions (e.g., clinician → participant, participant → participant)
• Speaker dominance indices
• Engagement rates between participants
• Transition matrices & dyadic graphs
• Temporal trends (with optional bins)
• Reliability: inter-coder sequence comparisons (e.g., Levenshtein distance)
• Correlation with treatment outcomes (e.g., ACOM, WAB) for longitudinal studies
• Turn quality (marker proportions for repairs/overlaps)
• Consistency over time (bin-level variability)
• Interaction structure (flow directionality between speakers)
• Individual engagement (relative contributions across sessions)
• Balance of participation/distribution metrics (e.g., Gini index, entropy, clinician–participant ratios)

---

Limitations

• Turn Overlap: current system assumes sequentialization - not uncommonly violated in group settings.
• Subjectivity: coder judgment needed for speaker boundaries and substantiality. Calibration recommended.
• Binary turn length: mark1 vs. mark2 is coarse; future versions may refine scale.
• Scalability: currently designed for up to 9 participants, future work could accommodate codes like C,P10, P11.

---

Profile of Word Errors and Retrieval in Speech (POWERS) coding

Measures

The POWERS coding system addresses the need to assess language abilities (particularly lexical retrieval) in conversation for people with aphasia. TAALCR facilitates quantification of the following subset of POWERS variables for both the clinician and client (see the POWERS manual for full details):

• filled pauses - disfluencies like "um", "uh", "er", etc.
• speech units - these more or less map onto non-punctuation tokens excluding filled pauses
• content words - nouns (including proper nouns), non-auxiliary verbs, adjectives, -ly-terminal adverbs, and numerals
• nouns - a subset of content words
• number of turns - a verbal contribution to the conversation with three types:
  • substantial turn - contains at least one content word
  • minimal turn - hands the turn back to the other conversation partner
  • subminimal turn (a nonce, non-canonical term) - not classifiable as either type above
• collaborative repair - sequences of turns devoted to overcoming communicative error/difficulty

Automation (reliability details pending)

TAALCR automates as much as possible. Below are descriptions of automatability and ICC2 utterance-level reliability metrics on a stratified (by study site, mild/severe aphasia profile, and pre-/post-tx test) random selection of XX samples (XX utterances).

• fully automated with regex and spaCy (en_core_web_trf):
  • filled pauses:
  • speech units:
  • content words:
  • noun count:
• semi-automated with a computational first pass followed by manual checks:
  • turn type:
• fully manual given the rich contextual dependencies:
  • collaborative repair

Typical Workflow

1. Tabularize utterances (if needed)
   If *Utterances*.xlsx files aren’t present, TAALCR will call RASCAL to read .cha files and tabularize utterances, assigning samples unique identifiers at the utterance and transcript levels.

2. Prepare POWERS coding files
   taalcr powers make creates full dataset plus reliability coding workbooks, with most coding automated.

3. Human coding
   Coders complete POWERS annotations in the generated spreadsheets.

4. Analyze
   taalcr powers analyze aggregates and reports POWERS metrics at the turn, speaker, and dialog levels.

5. Reliability evaluation
   taalcr powers evaulate matches reliability files and runs ICC2 evaluation.

6. Reliability subset (optional)
   taalcr powers reselect eselects reliability coding subset if ICC2 measures fail to meet threshold (0.7 a typical minimum).

Pipeline Commands

Command	Function (name)	Input	Output
powers make	Prepare POWERS coding files (make_POWERS_coding_files)	Either .cha files or utterance tables generated with RASCAL	POWERS coding spreadsheets for coders
powers analyze	Analyze POWERS coding (analyze_POWERS_coding)	Completed POWERS spreadsheets	Turn-, speaker-, and dialog-level aggregates
powers evaluate	Evaluate POWERS reliability (match_reliability_files, analyze_POWERS_coding)	Coder 2 + Coder 3 spreadsheets	Reliability metrics (ICC2, kappa, etc.)
powers reselect	Reselect POWERS reliability (reselect_POWERS_reliability)	Original + reliability spreadsheets	New reliability subset(s) for reassignment

---

Automation Validation

TAALCR includes CLI utilities to validate automatic POWERS coding against manual coding.

This workflow has two main steps:

1. Select Validation Samples

Use (stratified) random sampling to create a balanced subset of samples for manual validation.

Arguments:

• --stratify: Optional fields to group by (comma, space, or repeated flags) in random sample selection.

  Example: --stratify site,test or --stratify site --stratify test.

• --strata: Number of samples to draw per stratum (default: 5).

• --seed: Random number generator seed for reproducibility (default: 42).

Output:

• An Excel file POWERS_validation_selection_<timestamp>.xlsx containing the selected samples.

• The stratum_no column facilitates "chunking" the reliability subset. For example:

  • Code through stratum numbers 1 & 2
  • Evaluate reliability
  • Work through further strata if agreement is poor

• If POWERS coding tables exist in the input folder, labeled versions with stratum_no will also be written.

# Example
taalcr powers select \
  --stratify site,test \
  --strata 5 \
  --seed 42

2. Validate Automation

Merge the automatic and manual coding files for side-by-side comparison and reliability checks.

Requirements:

• Place your coding files in two subdirectories under the input folder:

  • Auto/ containing automatically generated coding files

  • Manual/ containing manually coded files

Arguments:

• --selection: Path to the selection Excel file from the previous step. Required if stratum_no is not already in the Manual coding files.

• --numbers: Optional comma- or space-separated list of stratum numbers to include (e.g., --numbers 1,2).

Output:

• An Excel file POWERS_Coding_Auto_vs_Manual.xlsx inside a new AutomationValidation/ folder. This file contains paired automatic and manual codes, restricted to the requested strata if specified.

# Example
taalcr powers validate \
  --selection taalcr_powers_select_output_250930/POWERS_validation_selection_250930_1530.xlsx \
  --numbers 1,2

Typical Workflow

1. Run powers select to generate a stratified subset of samples.

2. Manually code samples marked with stratum_no.

3. After manual coding, run powers validate to merge auto vs manual annotations.

4. Use the merged file to compute inter-coder reliability or other evaluation metrics.

---

🧪 Testing

This project uses pytest for its testing suite.
All tests are located under the tests/ directory, organized by module/function.

Running Tests

To run the full suite:

pytest

Run "quietly":

pytest -q

Run a specific test file:

pytest tests/test_samples/test_digital_convo_turns_analyzer.py

---

Status and Contact

I warmly welcome feedback, feature suggestions, or bug reports. Feel free to reach out by:

• Submitting an issue through the GitHub Issues tab

• Emailing me directly at: nsm [at] temple.edu

Thanks for your interest and collaboration!

(This project was previously developed under the working name “DIAAD”; the rename reflects a clarified scope and focus.)

---

Citation & Acknowledgments

Full details of the POWERS coding system can be found in the manual:

Herbert, R., Best, W., Hickin, J., Howard, D., & Osborne, F. (2013). Powers: Profile of word errors and retrieval in speech: An assessment tool for use with people with communication impairment. CQUniversity.

If TAALCR supports your work, please cite the repo:

McCloskey N. (2025). TAALCR: Toolkit for Aggregate Analysis of Language in Conversation, for Research. GitHub repository. https://github.com/nmccloskey/taalcr

---