From Long-Term Dialogue Memory to Personality Profiling: A Research Trajectory of Repeated Contraction Open Access
Zhao, Chunyao (Spring 2026)
Abstract
This thesis began with a broad question about long-term memory in dialogue: in long-form multi-party conversation, what kinds of information should be remembered, and how should they be represented so that they can support later reuse? Using Friends as the main corpus, the project initially aimed to move toward a memory-oriented system for long-horizon dialogue. However, as the work developed, that broader agenda repeatedly failed to stabilize. The project therefore evolved through a sequence of increasingly constrained stages rather than a single linear pipeline.
The early stages explored cross-episode reuse as a possible criterion for memory worthiness, then shifted to schema selection for dialogue information. Existing event and relation extraction frameworks provided useful starting points, but they proved insufficient for the kinds of socially and narratively specific information required in Friends. This led to a gradual contraction of the project toward a smaller but more workable representation: utterance-level events rewritten as self-contained natural-language event descriptions.
In the final stage, this surviving representation was reused in a personality profiling experiment. The hypothesis was that dialogue-derived Big Five profiles (profiling over the five dimensions of Agreeableness, Conscientiousness, Extraversion, Openness, and Neuroticism) would show a directional trend over time, especially within a season and between Season 1 and Season 10. Two scoring routes were implemented: subscene-level scoring followed by within-episode aggregation, and direct full-episode scoring. Neither method supported the expected monotonic trend. In both cases, the resulting curves fluctuated sharply from episode to episode, and the hypothesis was therefore not supported.
Follow-up comparisons showed that the two methods introduced different distortions: subscene-based scoring preserved local evidence but flattened it through averaging, while full-episode scoring often imposed a dominant evaluative direction and suppressed counter-evidence. The thesis does not claim to have built or disproved a general long-term memory system. Its contribution is instead a documented research trajectory: an originally broad memory-oriented project that contracted into a narrower, testable profiling experiment, whose final hypothesis was falsified, and whose intermediate representational results remained the most defensible outcome of the work.
Table of Contents
1 Introduction 1
1.1 Research Question 2
1.2 Thesis Statement 2
2 Background 4
2.1 Related Work 4
2.1.1 Long-term memory for large language models and dialogue agents 4
2.1.2 Schema-based representation of dialogue information 7
2.1.3 Personality profiling from dialogue 7
2.2 Why schema selection became the first obstacle 8
2.3 Surveying the existing schema landscape 9
2.4 Building the type tables: from literature review to empirical evidence 10
2.5 Closed schema as the default, and why it was insufficient 18
2.6 From “not enough types” to “tuple itself is not enough” 20
2.7 From open schema to a hybrid design 21
2.8 Why previous work was almost entirely tuple-based 23
2.9 Why natural-language-based memory became feasible now 25
3 Approach 27
3.1 Motivation and Initial Framing 27
3.2 Why Friends Was Chosen as the Long-Conversation Domain 28
3.3 The Earliest Approach: Direct Natural-Language Reuse Detection 29
3.4 The First Working Pipeline 30
3.4.1 Extracting candidate information from Episode 1 31
3.4.2 Reuse judgment in Episode 2 31
3.4.3 Block matching and evidence grounding 33
3.4.4 Failed candidate pruning with embeddings 34
3.4.5 Verification and final output 34
3.5 Cost, Scale, and the O(n²) Barrier 35
3.6 The Questions That Broke the First Route 37
3.7 The Memory Bank Proposal at the End of the Semester 38
3.8 What I Failed to Understand at the Time 41
3.9 Chapter Summary 42
3.10 Replicating the Life Event Extraction Task with GPT 43
3.10.1 Why Life Event Dialog Was Chosen 43
3.10.2 Stage I: The Single-Prompt Attempt 45
3.10.3 Stage II: Separating Recall from Precision 46
3.10.4 Stage III: Improving Precision by Revising the Evaluation Interface 47
3.10.5 Grammatical Errors and Internal Contradictions in the Gold Standard–Preposition Handling in the Gold Standard 49
3.10.6 Scope and Explicit Limitations of the Replication 60
3.10.7 What This Stage Actually Achieved 61
3.11 Applying the Life Event Schema to Friends and the Collapse of the Filtering Assumption 61
3.11.1 From Benchmark Replication to Corpus Extraction 61
3.11.2 Retaining the Filtered-Out Candidates 62
3.11.3 Linguistic Reclassification of Filtered Tuples and Information Loss 63
3.11.4 The Project-Level Meaning of This Shift 68
3.11.5 Concluding Remark for This Stage 69
3.12 From Memory System to Character Profiling 69
3.13 Stage I: Broadcasting Event Tuples into Character-Centered Data 71
3.14 Representation as a Character-Centered Node Graph 74
3.15 When a Correct Graph Still Fails to Produce Insight 77
3.16 The Role-Assignment Detour and Its Abrupt Interruption 82
3.17 Stage II: Hierarchical Merging as the First Serious Compression Strategy 83
3.18 Why the Merge Criteria Were Defined Backwards 86
3.19 The First Profiling Plan and Why It Failed 88
3.20 Action Label Distribution: A Real Output Without a Persuasive Pattern 89
3.21 Rebuilding Merging After the Collapse of the Hierarchical Plan 93
3.22 The First Redesign Still Tried to Do Too Much: Merge, Reuse, and Delete in One Pass 95
3.23 Candidate generation by character overlap 97
3.24 Restricting the Scope: Only Merge First, and Only Where the Standard Can Be Audited 99
3.25 Adding Self-Contained Natural-Language Event Descriptions 102
3.26 Extending the Same-Occurrence Standard Beyond the Utterance 104
3.27 When a Clear Standard Produces Almost No Compression 106
3.28 Cross-Episode Reuse and the Point at Which Linking Was Abandoned 107
3.29 From Event Representation to Personality Profiling 109
3.30 Choosing a Personality Profiling Schema 110
3.30.1 CoPE initially appeared to be a good fit 110
3.30.2 Why CoPE became impractical in the thesis setting 111
3.30.3 The final choice was FriendsPersona 112
3.31 From the Original FriendsPersona Pipeline to the Adapted Pipeline in This Thesis 113
3.32 Re-operationalizing MainSpeakerFinder for Profiling 115
3.32.1 What MSF is actually doing 115
3.32.2 Why a literal point-peak reproduction immediately failed 115
3.32.3 What the speaker curves revealed 116
3.32.4 The final thresholded peak-region interpretation 119
3.32.5 Why exact reproduction against the released FriendsPersona extraction was not possible 120
3.33 LLM-Based Personality Annotation Setup 120
4 Experiments 124
4.1 Testing the Hypothesis of Monotonic Personality Change 124
4.1.1 Hypothesis 124
4.1.2 First Test: Subscene-Based Aggregation 125
4.1.3 Second Test: Full-Episode Scoring 128
4.1.4 Conclusion: The Hypothesis Was Not Supported 129
5 Analysis 130
5.1 Follow-Up Comparisons 130
5.1.1 Subscene vs. Episode 130
5.1.2 Episode vs. Scene 136
5.1.3 Utterance vs. Event 137
5.2 What This Stage Established 138
6 Conclusion 141
6.1 Future Work 144
6.1.1 Immediate extensions: completing the personality profiling line 144
6.1.2 Medium-range expansion: other profiling tasks 145
6.1.3 Long-term direction: returning toward the memory question 145
A Appendix 147
A.1 Short-line/state-fact extraction prompt 147
A.2 Long-line/transformation-arc extraction prompt 148
B. Episode 1–Episode 2 reuse prompt 149
C.1 Single-prompt attempt 151
C.2 Rule-expansion prompt 151
D. Two-prompt decomposition 152
D.1 Recall prompt 153
D.2 Precision prompt 154
D.3 Lemmatize verb, move preposition, and resolve pronoun prompt 157
E.1 Classification Table 158
F. Broadcast table of Episode 1 (excerpt) 161
G. Annotated reuse pairs 162
H. Joint merge / reuse / delete prompt 168
I. Final merge prompt 170
J. Description prompt 173
K. Linking prompt 174
L. Big 5 scoring prompt 177
M. Dialog examples from the LED gold standard 180
Bibliography 184
Non-Printed Sources 187
List of Figures
3.1 fact list - facts 32
3.2 fact list - arcs 33
3.3 Reference Answer 35
3.4 cost estimate 36
3.5 memory bank 39
3.6 FE node 40
3.7 connecting 75
3.8 inferencing 76
3.9 character-event network 78
3.10 character-event zoomed in 79
3.11 single node in character-event network 80
3.12 info on a single node in character-event network 81
3.13 utterance-level event 85
3.14 scene-level event 85
3.15 episode-level event 86
3.16 utterance-level event 101
3.17 utterance-level event structure 102
3.18 utterance-level event description 104
3.19 peak plot 1 117
3.20 peak plot 2 117
3.21 peak plot 3 117
3.22 peak plot 4 118
3.23 peak plot 5 118
3.24 peak plot 6 118
3.25 anonymized event-based sub-scene input example 122
4.1 Rachel’s Season 1 Conscientiousness curve under subscene-based aggregation 126
4.2 other curves under subscene-based aggregation1 126
4.3 other curves under subscene-based aggregation2 127
4.4 other curves under subscene-based aggregation3 127
4.5 Rachel’s Season 1 Conscientiousness curve under full-episode scoring 129
5.1 global difference distribution 131
5.2 absolute difference by trait 133
5.3 aggregate by scene 137
List of Tables
2.1 Event Types Table 11
2.2 Relation Types Table 13
3.1 LED scores 49
3.2 our scores 49
3.3 action-category counts for Episodes 1 91
3.4 action-category counts for Episodes 2 92
5.1 overall comparison between subscene-mean scores and full-episode scores 131
5.2 trait-level comparison summary 132
5.3 correlation by trait 132
5.4 same-sign agreement by trait 133
5.5 top discrepancy cases 139
5.6 manual comparison summary for utterance vs. event 139
A.1 Classification Table 159
A.2 Broadcast table of Episode 1 (excerpt; first 30 rows) 161
A.3 Annotated reuse pairs 163
About this Master's Thesis
| School | |
|---|---|
| Department | |
| Degree | |
| Submission | |
| Language |
|
| Research Field | |
| Keyword | |
| Committee Chair / Thesis Advisor | |
| Committee Members |
Primary PDF
| Thumbnail | Title | Date Uploaded | Actions |
|---|---|---|---|
|
|
From Long-Term Dialogue Memory to Personality Profiling: A Research Trajectory of Repeated Contraction () | 2026-04-03 15:46:43 -0400 |
|
Supplemental Files
| Thumbnail | Title | Date Uploaded | Actions |
|---|