Master PluriTAL - Who's the werewolf - Speech comprehension and role prediction in the game of Werewolf

Aug 26, 2026

Uploaded by Logic

Speech comprehension and role prediction for the game of Werewolf is built to identify which players are werewolves, supporting a future AI player. The model learns from a corpus transcribed from the Chinese TV show Pandakill: 12 participants split into four werewolves, four ordinary villagers, and four divine villagers. Assuming the AI acts as an ordinary villager, the approach analyzes each reaction turn by intent classification of each sentence and rule-based summarization of meaning and attitude, combined with behavior analysis, to estimate the most probable werewolf team.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages70

Reading Time2 h 56 min

AccessFree to download

Introduction

Project design

State of the Art

Werewolf game rules

Roles

Mechanics

Speech processing

Data exploration

Corpus and annotation

Statistics

Intent classification

Without contextual features

With contextual features

Comparing the results

From classification to comprehension

Who attacks whom

Speech turn summary

Behavior processing

Behavior record

Who's the werewolf

Results analysis

Discussion

Appendices

Tools and libraries

Glossary of Werewolf game-specific expressions

Master PluriTAL Who's the werewolf Speech comprehension and role prediction in the game of Werewolf Author: Manying Zhang Supervisor: Dr. Kim Gerdes September 2018 Abstract The project is to build a model to predict the identities of the werewolves in the game of The Werewolves of Millers Hollow, aimed for future incorporation in an AI player program. The corpus data comes from the transcription of Chinese TV show Pandakill where 12 players are divided into 4 werewolves, 4 ordinary villagers and 4 divine villagers. The game has several speech and reaction turns until all the werewolves are eliminated, or either all the ordinary villagers or all the divine villagers are eliminated by werewolves. We assume that our AI plays as an ordinary villager to analyze all the speeches and behaviors before its reaction. The speech analysis is composed of 2 steps: the classiﬁcation of the intent of a single sentence, and the summarizing of sentence meaning and player attitude in a speech turn. The ﬁrst step is implemented with one of several machine learning and deep learning methods such as Multinomial naive Bayes, logistic regression, Linear support vector machine and Long Short Term Memory model in recurrent neural network; those methods are compared to each other. The second step, along with the behavior analysis part, is built by hand-crafted rules, which in the end gives a most probable Werewolf Team for each game turn. Acknowledgements I would ﬁrst like to thank my thesis advisor M. Kim Gerdes and my teacher M. Sylvain Kahane for their helpful guidance and advice. They gave me complete autonomy in choosing a subject that I love and always showed me an open door when I wanted to discuss the project. Besides, I would like also to thank my friends and my family, in particular Seginus for his endless loving support and his helpful comments and mathematical suggestions. Contents 1 Introduction .............................. 5 1.1 Project design ........................ 6 1.2 State of the Art ....................... 7 1.3 Werewolf game rules ..................... 9 1.3.a Roles ........................ 10 1.3.b Mechanics ..................... 11 2 Speech processing ........................... 14 2.1 Data exploration ....................... 14 2.1.a Corpus and annotation .............. 14 2.1.b Statistics ...................... 18 2.2 Intent classiﬁcation ...................... 20 2.2.a Without contextual features ........... 20 2.2.a.i Tf-idf ................... 21 2.2.a.ii Naive Bayes ............... 23 2.2.a.iii Logistic regression ............ 26 2.2.a.iv Support vector machine ......... 29 2.2.b With contextual features ............. 33 2.2.b.i word embedding ............. 34 2.2.b.ii Recurrent neural networks ....... 36 2.2.c Comparing the results ............... 39 2.3 From classiﬁcation to comprehension ............ 40 2.3.a Who attacks whom ................ 41 2.3.b Speech turn summary ............... 46 3 Behavior processing ......................... 49 3.1 Behavior record ........................ 49 3.2 Who's the werewolf ..................... 54 3.3 Results analysis ........................ 57 4 Discussion ............................... 61 Appendices 63 A Tools and libraries .......................... 63 B Glossary of Werewolf game-speciﬁc expressions .......... 63 1 Introduction The Werewolves of Millers Hollow (French: Les Loups-arous de Thiercelieux) 1 is a role-playing game created by French authors Philippe des Pallières and Hervé Marly that can be played with 8 to 47 players. The game is based on the Russian game Maﬁa2. It was nominated for the 2003 Spiel des Jahres award. There are numerous variants and extensions of the Werewolves game. Since its introduction in China, this game gained a lot of traction there. The unique entertainment and television system in the country has been a fertile ground for many television programs a

FAQs

What is the main goal of the project?

Predict the identities of the werewolves in Werewolf so it can be integrated into an AI player program.

What data source is used to train the model?

A transcription of the Chinese TV show Pandakill, containing gameplay from 12 players divided into werewolves, ordinary villagers, and divine villagers.

How is speech analysis performed within each game turn?

It uses two steps: intent classification for each sentence, then rule-based summarization of sentence meaning and the player’s attitude, combined with behavior analysis.

Aug 26, 2026

Uploaded by Logic