NarraScore: Narrative Proxy & Soundtrack Generation

• NarraScore is a term for two distinct computational approaches: one as a proxy for narrative quality via predicted upvotes and one as a framework for narrative-conditioned soundtrack synthesis.
• The narrative quality usage segments text with CNNs and recurrent gating models to predict log-upvotes, effectively modeling inter-region dependencies in social media narratives.
• In soundtrack synthesis, NarraScore employs affective control using Valence-Arousal trajectories and dual-branch injection to condition pre-trained music generators for coherent audio output.

Searching arXiv for “NarraScore” to verify relevant papers and terminology. NarraScore denotes two distinct research usages centered on computational treatments of narrative structure. In Editor's term usage derived from "Predicting the Quality of Short Narratives from Social Media," it refers to a proxy narrative-quality score given by predicted upvotes for short narratives, with the target modeled as $\mu_i = E[\ln y_i]$ and the test-time score recovered as $\hat{y}_n = \exp(\hat{\mu}_n)$ (Wang et al., 2017). In "NarraScore: Bridging Visual Narrative and Musical Dynamics via Hierarchical Affective Control," the term names a hierarchical framework for long-form video soundtrack synthesis that uses frozen Vision-LLMs as affective sensors, extracts Valence-Arousal trajectories, and conditions a frozen MusicGen backbone through a Dual-Branch Injection strategy (Wen et al., 9 Feb 2026). The two usages are methodologically different, but both treat compressed surrogate signals as operational handles on otherwise difficult narrative phenomena.

1. Terminological scope and disambiguation

The term has no single canonical meaning across the supplied literature. In the 2017 work, the paper "does not introduce a separate name for the score beyond ‘predicted upvotes,’" and the summary explicitly states that one may think of the target as a continuous quality score (Wang et al., 2017). In the 2026 work, NarraScore is the formal name of a generative system for soundtrack synthesis (Wen et al., 9 Feb 2026).

Usage	Object	Core signal
Editor's term “NarraScore”	Predicted quality of short narratives	Predicted upvotes, trained on $\ln y$
NarraScore	Long-form video soundtrack generation	Valence-Arousal trajectories plus global semantic summary

This distinction matters because the 2017 usage is an evaluation proxy for narratives, whereas the 2026 usage is an architecture for narrative-conditioned generation. A common misconception would be to treat NarraScore as a single benchmark or a single metric; the supplied literature instead shows two separate research programs sharing a name.

2. NarraScore as proxy narrative quality prediction

In the 2017 study, the target variable is defined as the number of upvotes received by story $i$, $y_i$, on Quora (Wang et al., 2017). Because the empirical distribution of $y$ is highly skewed, described as "roughly Pareto," the model is trained to predict $\mu_i = E[\ln y_i]$ and minimizes

$L(\theta) = \sum_n (\ln y_n - \hat{\mu}_n)^2$

where $\hat{\mu}_n = f(x_n; \theta)$ is the network’s real-valued output given the narrative text $x_n$ and parameters $\hat{y}_n = \exp(\hat{\mu}_n)$0. At test time, the score is $\hat{y}_n = \exp(\hat{\mu}_n)$1 (Wang et al., 2017). The loss is also described as equivalent to maximum likelihood estimation under a constant-$\hat{y}_n = \exp(\hat{\mu}_n)$2 log-normal assumption.

The dataset construction is central to the definition of this score. The initial crawl contains 54,484 answers across 21 Quora topics. A set of 1,400 answers was manually labeled, using AMT and an active-learning loop, as story versus non-story; a random-forest classifier with BoW features and $\hat{y}_n = \exp(\hat{\mu}_n)$3 then labeled all answers. After filtering out texts with fewer than 50 words and answers with fewer than 50 views, the resulting corpus contains 28,320 stories under 742 questions, split into 21,230 training examples, 2,832 development examples, and 4,258 test examples (Wang et al., 2017).

The paper characterizes this as, to the authors’ knowledge, "the first large-scale study for automatic evaluation of narrative quality" (Wang et al., 2017). At the same time, it explicitly treats upvotes as an approximate proxy rather than a direct measure of intrinsic narrative quality. This limitation is substantive rather than incidental: the score operationalizes public response to social-media narratives, not a ground-truth aesthetic or narratological judgment.

3. Regional, sequential, and holistic readers for short narratives

The 2017 methodology represents the full question-answer pair as multiple textual regions. The answer is chopped into $\hat{y}_n = \exp(\hat{\mu}_n)$4 regions of fixed size $\hat{y}_n = \exp(\hat{\mu}_n)$5 words, plus one special region holding the question. Each region is represented as a $\hat{y}_n = \exp(\hat{\mu}_n)$6 matrix of pre-trained word2vec embeddings. A 3-layer CNN with shared filters across answer regions maps each block to a $\hat{y}_n = \exp(\hat{\mu}_n)$7-dimensional regional embedding $\hat{y}_n = \exp(\hat{\mu}_n)$8, while the question region uses an otherwise identical CNN with its own filter parameters to produce $\hat{y}_n = \exp(\hat{\mu}_n)$9 (Wang et al., 2017).

In the simplest regional model, the document representation is

$\ln y$0

and this vector is passed through two fully connected ReLU layers of size 128 before a final linear output $\ln y$1 (Wang et al., 2017). The architecture is designed to model local narrative segments without collapsing the entire text into a single sequence representation from the outset.

The sequential reader introduces recurrent interdependence among regions. Its recurrent state is updated as

$\ln y$2

with a gate

$\ln y$3

and gated state

$\ln y$4

The concatenated vector $\ln y$5 is again fed to two 128-unit ReLU layers and then to the linear predictor (Wang et al., 2017). Both a simple RNN version and an LSTM variant were tested.

The holistic reader keeps the same recurrent backbone but changes the gate so that each $\ln y$6 can access all region embeddings and all hidden states:

$\ln y$7

with

$\ln y$8

This design lets each region be modulated with full-document context rather than only left-to-right context (Wang et al., 2017). The summary states that "the LSTM + holistic gate was strongest," which is consistent with the paper’s emphasis on explicitly modeling inter-region dependencies.

The models also incorporate 26 hand-crafted linguistic features, including counts from Harvard General Inquirer categories such as OBJECT, POSITIVE, and NEGATIVE, together with $\ln y$9 views, number of images, number of words, and a simple grammar-error count. These features are concatenated into the first fully connected layer in all neural models (Wang et al., 2017).

4. Training configuration, baselines, and empirical behavior in narrative scoring

The 2017 models use 300-dimensional word2vec embeddings trained on approximately 4 GB of fiction and plot summaries, held fixed during training. The CNN configuration is specified as three convolutional layers: 32 filters of size $i$0 with stride 3, 32 filters of size $i$1 with stride 2, and 16 filters of size $i$2 with stride 1, each followed by ReLU and $i$3 max-pooling. The region embedding dimension is $i$4, dropout is 50% after the CNN and between each fully connected layer, the region count is $i$5, the region size is $i$6 words, and the question occupies region 0, for a total of 11 regions. The RNN hidden size is 100, gradient clipping is 1.0, and training runs for 100 epochs with hyperparameters selected on the development set (Wang et al., 2017).

Evaluation is performed using mean squared error on $i$7, reported as MSE of predictions versus $i$8 upvotes. The baselines include a random-forest regressor with 1,000 trees using BoW plus meta-features, Lasso with $i$9, SVM-RBF, and plain RNN and LSTM document models using average-pooling over hidden states before the same two-layer 128-unit head (Wang et al., 2017).

Model	MSE
Random Forest baseline	0.5419
RNN-only encoder	0.4895
LSTM-only encoder	0.4776
Regional reader	0.4621
Sequential reader (RNN-gated)	0.4442
Sequential + LSTM gating	0.4506
Holistic reader + RNN	0.4472
Holistic reader + LSTM	0.4438

The reported improvements are stated relative to the random-forest baseline: approximately 9.7% for the RNN-only encoder, 11.9% for LSTM-only, 14.7% for the regional reader, 18.0% for the sequential reader, 17.5% for holistic reader plus RNN, and 18.1% for holistic reader plus LSTM (Wang et al., 2017). The ablations are also specific. Excluding all text and using only the 26 meta-features gives MSE $y_i$0, which is worse than the random forest. Adding only the answer text to the holistic network gives MSE $y_i$1, and including the question text yields the full holistic plus LSTM result of 0.4438 (Wang et al., 2017).

The authors hypothesize that the question’s simpler and more uniform language is easier for the network to learn from than the more varied answer styles. They also identify a broader pattern: explicitly modeling inter-region dependencies through gating yields consistent gains (Wang et al., 2017). A plausible implication is that narrative evaluation benefits not only from lexical content but also from document-level dependency structure between narrative segments and their prompts.

5. NarraScore as hierarchical affective control for soundtrack generation

In the 2026 paper, NarraScore is a hierarchical framework for synthesizing coherent soundtracks for long-form videos (Wen et al., 9 Feb 2026). Its stated motivation is the conjunction of three obstacles: computational scalability, temporal coherence, and "semantic blindness to evolving narrative logic." The framework is predicated on the claim that emotion serves as a "high-density compression of narrative logic," specifically through a continuous Valence-Arousal representation (Wen et al., 9 Feb 2026).

The system has two stages, Perception and Synthesis, linked by a Dual-Branch Injection strategy. A frozen VLM processes sparse video frames sampled at 1 Hz and emits two outputs: a global semantic description $y_i$2, summarized in terms of genre, instrumentation, atmosphere, and pacing, and a local affective trajectory $y_i$3 tracking frame-level Valence-Arousal. A frozen MusicGen backbone then conditions on $y_i$4 through cross-attention, called the Global Semantic Anchor, and on a temporally super-resolved version of the affective trajectory, $y_i$5, through residual injection, called the Token-Level Affective Adapter (Wen et al., 9 Feb 2026).

The perception stage uses what the paper calls Semantically-Anchored Temporal Alignment and Instruction-Driven Semantic Steering. Frames $y_i$6 are sampled at 1 Hz, and textual "semantic clocks" $y_i$7 are inserted so that the VLM’s native text-vision attention handles temporality. A learned system prompt $y_i$8 is prepended in order to steer the backbone away from object enumeration and toward high-level affect reasoning, yielding an input sequence

$y_i$9

Latent affective probing is then performed by taking the final hidden layer $y$0, spatially averaging the patch embeddings $y$1 for frame $y$2 to get

$y$3

and projecting with an MLP into the Valence-Arousal plane, with clipping:

$y$4

This may be written compactly as $y$5 (Wen et al., 9 Feb 2026).

The synthesis stage separates global and local conditioning. The Global Semantic Anchor injects $y$6 into MusicGen’s cross-attention at every decoding step, anchoring overall style. The Token-Level Affective Adapter first interpolates sparse $y$7 to the acoustic token rate $y$8, then applies dilated convolutions $y$9:

$\mu_i = E[\ln y_i]$0

Residual injection is applied in shallow decoder layers, $\mu_i = E[\ln y_i]$1, according to

$\mu_i = E[\ln y_i]$2

where $\mu_i = E[\ln y_i]$3 is a learnable scalar initialized to 0 (Wen et al., 9 Feb 2026). The paper states that injecting only into early layers guides the generation toward the desired tension without disturbing deeper harmony synthesis.

6. Objectives, efficiency claims, and reported results of the 2026 framework

The 2026 NarraScore trains only small interface components rather than the backbones. The affective probe is trained with a hybrid $\mu_i = E[\ln y_i]$4 regression objective:

$\mu_i = E[\ln y_i]$5

MusicGen remains frozen; only $\mu_i = E[\ln y_i]$6 and $\mu_i = E[\ln y_i]$7 are trained using standard autoregressive cross-entropy,

$\mu_i = E[\ln y_i]$8

Regularization includes dropout of 0.1 in projection MLPs, zero initialization of $\mu_i = E[\ln y_i]$9, and a two-stage training schedule: 150 epochs for the projector and 50 epochs for the temporal adapter (Wen et al., 9 Feb 2026).

The computational-efficiency claims are also explicit. The framework uses no dense video-to-audio cross-attention at every frame, only 1 Hz sampling plus text clocks; no full architectural cloning or heavy adapters, only a small MLP probe and shallow residual injections; and sliding-window inference with overlap, using sparse keyframes for $L(\theta) = \sum_n (\ln y_n - \hat{\mu}_n)^2$0, windowed affective extraction for $L(\theta) = \sum_n (\ln y_n - \hat{\mu}_n)^2$1, and prompt prefixing of previously generated audio tokens for causal continuation (Wen et al., 9 Feb 2026). The paper states that memory growth stays near-linear in $L(\theta) = \sum_n (\ln y_n - \hat{\mu}_n)^2$2 and that the extra parameter count is negligible.

The experimental setup uses a continuous movie dataset with frame-level Valence-Arousal annotations totaling 884 minutes for video-to-emotion, and a music emotion dataset with per-second Valence-Arousal labels totaling 1,351 minutes for emotion-to-music. Preprocessing includes vocal separation with Demucs, 30-second clips with 15-second overlap, and discarding clips with more than 40% silence (Wen et al., 9 Feb 2026).

Objective evaluation uses FAD, FD, KLD, and ImageBind score IB. Subjective evaluation with 10 users uses Emotional Dynamic Consistency, Global Style Matching, Long-term Coherence, Music Quality, and Overall Preference (Wen et al., 9 Feb 2026).

Method	Objective summary	Subjective summary
GT	FAD 0.0, FD 0.0, KLD 0.0, IB 0.241	Not reported
VidMuse	FAD 2.459, FD 29.946, KLD 0.734, IB 0.202	EDC 2.29, GSM 2.49, LTC 2.49, MQ 2.13, OP 2.18
Video2Music	FAD 14.954, FD 115.008, KLD 1.269, IB 0.100	EDC 1.48, GSM 1.56, LTC 2.41, MQ 3.39, OP 1.88
GVMGen	FAD 2.362, FD 41.466, KLD 0.350, IB 0.213	EDC 1.65, GSM 1.69, LTC 1.41, MQ 1.64, OP 1.34
M2UGEN	FAD 9.647, FD 74.625, KLD 0.953, IB 0.182	Not reported
Caption2Music	Not reported	EDC 2.66, GSM 2.77, LTC 2.88, MQ 3.18, OP 2.82
NarraScore	FAD 1.923, FD 36.411, KLD 0.320, IB 0.219	EDC 2.86, GSM 3.02, LTC 3.15, MQ 3.41, OP 3.06

The ablations isolate the contributions of the affective and semantic branches. Removing the Narrative-Aware Affective Reasoning raises FAD from 1.923 to 3.009. Skipping the Holistic Musical Conceptualization loses global style stability. The best performance is reported when approximately 75% of shallow blocks receive affective injection (Wen et al., 9 Feb 2026). The qualitative summary states that baseline spectrograms are either overly flat, noisy, or erratic, while NarraScore balances clear harmonic foundations with rhythmic transients aligned to the narrative Valence-Arousal curve.

7. Comparative interpretation

Across the two usages, NarraScore indexes two different operationalizations of narrative structure. In the 2017 work, narrative quality is reduced to a scalar proxy derived from social response, and the core technical problem is how to map segmented text into a prediction of $L(\theta) = \sum_n (\ln y_n - \hat{\mu}_n)^2$3 upvotes (Wang et al., 2017). In the 2026 work, narrative logic is compressed into a continuous Valence-Arousal trajectory and a global semantic summary, and the technical problem is how to inject those signals into a pre-trained music generator while preserving coherence and scalability (Wen et al., 9 Feb 2026).

This suggests a broader methodological pattern rather than a single unified framework. In both cases, the narrative object itself is not modeled in full fidelity; instead, a compressed surrogate is used because direct supervision is difficult. In the earlier case, the surrogate is public endorsement on a crowd-powered website. In the later case, the surrogate is the evolving emotional arc of a video. The supplied literature therefore presents NarraScore not as a single stable concept, but as a name associated with computational strategies for rendering narrative phenomena tractable through proxy signals, structured representations, and neural conditioning mechanisms.