CoD-V: Video Complaint Text Generation

• CoD-V is a multimodal task that converts complaint videos into explicit text, integrating visual cues, reviews, and emotion labels for accurate grievance articulation.
• It is distinct from video captioning and summarization by focusing on defect details, affected aspects, and the conveyed emotional tone.
• The system leverages retrieval-augmented generation and emotion conditioning to improve metrics like ROUGE and BLEU, enhancing complaint retention.

Searching arXiv for the cited CoD-V and related complaint-understanding papers. Complaint Description from Videos (CoD-V) is a multimodal generation task that converts user-uploaded complaint videos into explicit complaint text intended to help users articulate grievances that are visually evident but only vaguely expressed in writing. In the formulation introduced in "When Words Can't Capture It All: Towards Video-Based User Complaint Text Generation with Multimodal Video Complaint Dataset" (Das et al., 24 Sep 2025), the input is a complaint video together with associated review context and an emotion label, and the output is a coherent complaint description that should accurately encapsulate the user’s complaint while encoding the user’s emotional state. CoD-V is therefore situated at the intersection of video-language modeling, complaint mining, retrieval-augmented generation, and affect-aware conditional generation, but it is explicitly distinguished from standard video captioning, video summarization, and complaint classification.

1. Task definition and conceptual scope

CoD-V is defined over complaint videos $V \in \mathbb{R}^{F \times 3 \times W \times H}$, corresponding textual reviews $R=\{r_1,r_2,\dots,r_m\}$, and an emotion label $$e \in \{\text{dissatisfaction}, \text{blame}, \text{frustration}, \text{disappointment}\}$$. The task is to learn a model $\mathcal{F}$ such that

$Y = \mathcal{F}(V, P, e),$

where $P$ denotes the prompt and $Y$ is the generated complaint description (Das et al., 24 Sep 2025). The review text is often vague, so the operational burden falls on the model to recover defect semantics, affected product aspects, and complaint tone from the video and auxiliary context.

The task is explicitly not standard video captioning. Captioning may produce generic visual descriptions such as object presence or observed actions, whereas CoD-V is intended to generate a complaint narrative. It is likewise not ordinary video summarization, because summarization can compress events without preserving complaint-relevant defect semantics or complaint tone. It is also not explainable complaint mining or complaint classification, both of which generally return labels, aspects, rationales, or causes rather than full complaint text (Das et al., 24 Sep 2025).

A central conceptual distinction in CoD-V is the difference between observable content and complaint articulation. A short review such as “worst product” paired with a video of a broken headphone or malfunctioning keyboard contains complaint evidence that text-only complaint mining systems cannot fully exploit. CoD-V therefore treats complaint generation as a user-assistance problem: the generated output should state the defect, the affected aspect, and, ideally, the user’s emotional stance clearly enough for customer-service resolution (Das et al., 24 Sep 2025). This suggests that CoD-V is not merely a generative reformulation of sentiment analysis, but a task in which pragmatic usefulness is part of the target representation.

2. ComVID dataset and label structure

The benchmark dataset for CoD-V is ComVID, described as a video complaint dataset containing 1,175 complaint videos, corresponding descriptions, and emotion labels (Das et al., 24 Sep 2025). Collection began from an initial pool of 1,200 Amazon review videos scraped using Beautiful Soup, with emphasis on 1-star and 2-star reviews. The collected metadata included review ID, rating, review text, aspects, domain name, product name, and associated m3u8 video links, which were then transcoded to mp4 format. After curation, the final dataset retained 1,175 samples (Das et al., 24 Sep 2025).

ComVID spans four domains. Electronics contains 655 reviews, Household 273, Fashion 202, and Others 45 (Das et al., 24 Sep 2025). The dataset supports analysis over eight complaint aspects, although the paper reports a somewhat inconsistent aspect inventory. From the text and tables, the aspect set appears to include combinations of Quality, Functionality, Defective or Defects, Missing items or Missing, Refunds, Performance, Delay, and Design (Das et al., 24 Sep 2025). Domain-specific examples are also reported: Fashion includes shoes, bags, watches, and T-shirts; Electronic Products include mouse, keyboard, headphone, and trimmer; Household includes bottle, plates, and plastic pots; Others includes tent and rain-coat (Das et al., 24 Sep 2025).

Each ComVID instance contains at least a complaint video, original review text, star rating, a manually written gold complaint description, and an emotion label (Das et al., 24 Sep 2025). The gold description is not treated as a short caption. It is a manually authored complaint narrative intended to capture the actual issue visible in the video. Representative targets illustrate a strongly template-like but semantically explicit style. One ground-truth mouse complaint is reported as: “The user wants to convey about the complaint of mouse. The user claims that the mouse is defective. He has received a damaged product. The scrolling button is not working properly.” A keyboard example states: “The user wants to convey the complaint that the keyboard is not working. The USB jack is also not working. Wrong typing occurs on the screen. words and trigger the escape button.” (Das et al., 24 Sep 2025)

The dataset description also records six samples with non-English review texts, specifically in Hindi, but no multilingual analysis is reported (Das et al., 24 Sep 2025). The paper does not report train, validation, and test split sizes, average video duration, average description length, per-aspect counts, or per-emotion counts. This absence is methodologically significant because it limits direct interpretability of generation difficulty and reproducibility.

3. Annotation process and emotional-state conditioning

ComVID was annotated in two phases by a team of five expert linguists: one doctoral-level annotator designated Category A, two master’s-level annotators designated Category B, and two undergraduate annotators designated Category C. All annotators were proficient in Hindi and English (Das et al., 24 Sep 2025). In Phase 1, the Category A annotator produced 50 gold-standard samples to establish reference style and quality. One Category B and one Category C annotator then wrote descriptions for half the dataset, while the remaining two annotators reviewed descriptions for accuracy, coherence, and guideline adherence. Cross-validation was performed on samples not originally annotated by a reviewer, and only samples receiving an acceptance tag were retained as gold-standard descriptions (Das et al., 24 Sep 2025).

Phase 2 covered emotion annotation. Category A and Category B annotators independently assigned emotion labels, and disagreements were resolved by consensus. The reported inter-annotator agreement is Fleiss’ $\kappa = 0.64$, described as substantial agreement. Annotators were paid $0.50 per sample (Das et al., 24 Sep 2025). The paper explicitly rejects fixed-length constraints on complaint descriptions, arguing that complaint articulation requires flexible length and structure to preserve context and information density.

Emotion is treated as an auxiliary conditioning signal rather than the primary prediction target. The emotion label set is $$\{\text{dissatisfaction}, \text{blame}, \text{frustration}, \text{disappointment}\}$$, and the model input is enriched with the user’s emotional state during generation (Das et al., 24 Sep 2025). The rationale is that a complaint description should preserve not only defect details but also complaint tone.

Empirically, emotion conditioning improves most reported generation metrics. For the proposed model, adding emotion increases ROUGE-1 from 0.55 to 0.59, ROUGE-2 from 0.40 to 0.47, ROUGE-L from 0.51 to 0.56, BLEU-1 from 0.65 to 0.69, BLEU-2 from 0.57 to 0.63, BLEU-L from 0.53 to 0.59, BERTScore from 0.91 to 0.93, METEOR from 0.46 to 0.51, and MoverScore from 0.14 to 0.24, while perplexity changes from 95.39 to 97.16 (Das et al., 24 Sep 2025). The paper interprets this as improved contextual relevance and emotional consistency. A plausible implication is that affect-aware conditioning helps the model avoid generating visually correct but pragmatically flat complaint descriptions.

4. Modeling architecture and retrieval-augmented generation

The principal CoD-V model is a multimodal Retrieval-Augmented Generation embedded VideoLLaMA2-7b system, also referred to as MR + VideoLLaMA2-7b (Das et al., 24 Sep 2025). The framework has two stages: multimodal retrieval and supervised fine-tuning for complaint generation.

In the retrieval stage, the corpus is constructed from a large filtered subset of a public Amazon product review dataset containing text and images. The source dataset has 75.26 million reviews, and filtering retains reviews with rating 1, text longer than 150 characters, and at least one image (Das et al., 24 Sep 2025). For each retrieved review, CLIP encodes text and images. The text embedding is

$$\mathbf{e}_t = \mathrm{TextEncoder}(T) \in \mathbb{R}^d,$$

and image embeddings are

$R=\{r_1,r_2,\dots,r_m\}$0

If a review contains multiple images, the image embeddings are averaged:

$R=\{r_1,r_2,\dots,r_m\}$1

The multimodal review embedding is then

$R=\{r_1,r_2,\dots,r_m\}$2

and the embedding matrix is indexed with FAISS using

$R=\{r_1,r_2,\dots,r_m\}$3

For a complaint video query, GMFlow is used to extract four keyframes, which are encoded by CLIP as $R=\{r_1,r_2,\dots,r_m\}$4. These are averaged to form

$R=\{r_1,r_2,\dots,r_m\}$5

The corresponding product aspects are also encoded textually, yielding $R=\{r_1,r_2,\dots,r_m\}$6, and the final query is

$R=\{r_1,r_2,\dots,r_m\}$7

Top-$R=\{r_1,r_2,\dots,r_m\}$8 nearest neighbors are retrieved from FAISS and used as context (Das et al., 24 Sep 2025).

In the generation stage, VideoLLaMA2-7b is fine-tuned on supervised video-text complaint pairs, and inference conditions on the complaint video, the user’s emotional state, and the retrieved complaint reviews (Das et al., 24 Sep 2025). The paper does not provide a detailed token-level fusion specification inside VideoLLaMA2-7b, an explicit generation loss, the values of $R=\{r_1,r_2,\dots,r_m\}$9 and $$e \in \{\text{dissatisfaction}, \text{blame}, \text{frustration}, \text{disappointment}\}$$0, the prompt template, or decoding parameters. It reports the following implementation details for VideoLLaMA2-7b and its LoRA fine-tuned variant: LoRA rank 64, LoRA alpha 128, batch size 16, learning rate $$e \in \{\text{dissatisfaction}, \text{blame}, \text{frustration}, \text{disappointment}\}$$1, multimodal projector learning rate $$e \in \{\text{dissatisfaction}, \text{blame}, \text{frustration}, \text{disappointment}\}$$2, cosine learning-rate scheduler, and training for 10 epochs (Das et al., 24 Sep 2025).

The central modeling claim is that retrieval does not supply generic world knowledge; it supplies semantically aligned negative product reviews with text and images, thereby providing defect phrasing, complaint structure, and domain-specific context that the base VLM may not otherwise recover (Das et al., 24 Sep 2025). The paper summarizes the contribution of multimodal retrieval augmentation as roughly 3–4% improvement over fine-tuned VLMs alone.

5. Evaluation protocol and Complaint Retention

The evaluation suite for CoD-V combines standard NLG metrics with a complaint-specific metric called Complaint Retention (CR). The standard metrics are BLEU, ROUGE, BERTScore, MoverScore, METEOR, Perplexity, Flesch Reading Ease Score, and Coleman-Liau Index Score (Das et al., 24 Sep 2025). BLEU and ROUGE are used for lexical overlap, BERTScore and MoverScore for semantic similarity, METEOR for precision-recall-oriented semantic and lexical alignment, Perplexity for fluency, and FRES and CLIS for readability and complexity.

The distinctive metric is CR, which is intended to measure whether the generated text preserves the complaint nature rather than merely producing a fluent or generic video description (Das et al., 24 Sep 2025). CR has three components. The first is the normalized Vader sentiment score. For sentence $$e \in \{\text{dissatisfaction}, \text{blame}, \text{frustration}, \text{disappointment}\}$$3,

$$e \in \{\text{dissatisfaction}, \text{blame}, \text{frustration}, \text{disappointment}\}$$4

and for sample $$e \in \{\text{dissatisfaction}, \text{blame}, \text{frustration}, \text{disappointment}\}$$5 with $$e \in \{\text{dissatisfaction}, \text{blame}, \text{frustration}, \text{disappointment}\}$$6 sentences,

$$e \in \{\text{dissatisfaction}, \text{blame}, \text{frustration}, \text{disappointment}\}$$7

The second is the Emotion Score, computed with Text2Emotion over Happy, Angry, Surprise, and Fear:

$$e \in \{\text{dissatisfaction}, \text{blame}, \text{frustration}, \text{disappointment}\}$$8

The third is the Aspect Score, where GPT-4 is prompted to determine whether each annotated aspect is present in the generated text:

$$e \in \{\text{dissatisfaction}, \text{blame}, \text{frustration}, \text{disappointment}\}$$9

Complaint Retention is then defined as the average of normalized sentiment, emotion score, and aspect score:

$\mathcal{F}$0

This formulation is used to distinguish CoD-V from summary generation and ordinary video description (Das et al., 24 Sep 2025).

The stated methodological argument is that standard generation metrics can reward outputs that are fluent and even semantically plausible while still omitting the actual complaint. A video summary may therefore be factually acceptable but operationally unhelpful for complaint resolution. CR is designed to reward preservation of complaint polarity, emotional intensity, and complaint aspects (Das et al., 24 Sep 2025). At the same time, the metric depends on Vader, Text2Emotion, and GPT-4-based aspect checking, so the paper itself exposes a notable evaluation dependency on external tools.

Human evaluation is also reported. On 80 randomly sampled test instances, Category B annotators rated outputs on 1–5 scales for Adequacy, Fluency, Emotion Encoding, Relevance, and Visual Content Alignment, with Category A cross-verification. The figure summary states that VideoLLaMA2-7b+MR outperformed BLIP-VQA-BASE and VideoLLaVA2-7b, especially on emotion encoding and nuanced emotional context capture (Das et al., 24 Sep 2025).

6. Empirical findings, failure modes, and relation to adjacent complaint-understanding work

Among the compared systems, the strongest overall model is VideoLLaMA2-7b+MR in the emotion setting. Its reported scores are ROUGE-1 0.59, ROUGE-2 0.47, ROUGE-L 0.56, BLEU-1 0.69, BLEU-2 0.63, BLEU-L 0.59, BERTScore 0.93, FRES 79.58, CLRS or CLIS 7.27, Perplexity 97.16, METEOR 0.51, and MoverScore 0.24 (Das et al., 24 Sep 2025). A strong comparator, VideoLLaVA2-7b+MR with emotion, achieves ROUGE-1 0.57, ROUGE-2 0.46, ROUGE-L 0.54, BLEU-1 0.68, BLEU-2 0.61, BLEU-L 0.58, BERTScore 0.92, METEOR 0.51, and MoverScore 0.21 (Das et al., 24 Sep 2025). Without emotion, VideoLLaMA2-7b+MR still leads, with ROUGE-1 0.55, ROUGE-2 0.40, ROUGE-L 0.51, BLEU-1 0.65, BLEU-2 0.57, METEOR 0.46, and MoverScore 0.14 (Das et al., 24 Sep 2025).

A separate zero-shot comparison across Summary Generation, Video Description, and CoD-V is used to argue that CoD-V is a distinct task. For the proposed model, the CR values are 0.59 for summary generation, 0.58 for video description, and 0.62 for CoD-V (Das et al., 24 Sep 2025). Under CoD-V specifically, the reported values are ROUGE-1 0.39, ROUGE-2 0.27, ROUGE-L 0.36, BERTScore 0.89, METEOR 0.34, Vader score $\mathcal{F}$1, and CR 0.62 (Das et al., 24 Sep 2025). The negative Vader score is treated as evidence that complaint generation preserves negative complaint polarity more effectively than ordinary summary or description formulations.

Qualitative analysis reveals both strengths and characteristic failure modes. In the defective mouse example, the generated output from VideoLLaMA2-7b+MR with emotion closely matches the ground truth, preserving product identity, defectiveness, damage, specific functional failure, and complaint tone (Das et al., 24 Sep 2025). In the keyboard example, the model correctly identifies malfunction and wrong typing but hallucinates inability to contact customer care, a refund issue, and explicit blame toward Amazon, none of which appear in the ground truth (Das et al., 24 Sep 2025). The paper therefore associates retrieval augmentation with two opposed effects: improved grounding on average and a tendency to inject plausible but unsupported complaint boilerplate. Other visible failure modes include missing defect details, generic wording, weak emotional grounding in emotion-agnostic variants, and semantic drift from observed malfunction to broader dissatisfaction narratives (Das et al., 24 Sep 2025).

The immediate research context of CoD-V includes adjacent complaint-understanding systems that do not themselves solve free-form complaint description generation. "Deciphering the complaint aspects: Towards an aspect-based complaint identification model with video complaint dataset in finance" (Das et al., 26 Feb 2025) studies multimodal aspect-based complaint identification in financial videos rather than complaint description generation. Its task predicts, for each of five financial aspects, whether the aspect is absent, present as non-complaint, or present as complaint, using speech-derived text and video frames as aligned multimodal evidence (Das et al., 26 Feb 2025). This work is relevant because it shows how complaint-bearing content in videos can be localized and structured before any generative stage. "Talk, Snap, Complain: Validation-Aware Multimodal Expert Framework for Fine-Grained Customer Grievances" (Singh et al., 18 Nov 2025) is closer to the semantic decomposition problem: it predicts complaint aspect category detection and severity detection from multimodal customer-support dialogues and aligned visual evidence, using routed Chain-of-Thought experts, a Semantic Alignment Score, and a validation mixture-of-experts layer (Singh et al., 18 Nov 2025). That framework still stops short of free-form complaint generation, but it supplies a plausible architecture pattern for multimodal complaint reasoning under evidence validation.

The main limitations of CoD-V as currently instantiated are clearly visible in the benchmark design. ComVID contains only 1,175 videos, is heavily skewed toward electronics, is restricted to Amazon review data, and uses coarse emotion labels (Das et al., 24 Sep 2025). The retrieval corpus may bias generation toward e-commerce-specific narratives such as refunds and platform-specific complaint phrasing. The aspect inventory is reported inconsistently, split details are omitted, and CR depends on external tools and a GPT-4-based aspect checker (Das et al., 24 Sep 2025). These constraints do not invalidate the task formulation, but they delimit the present benchmark’s scope. A plausible implication is that the most robust future CoD-V systems will need larger and more diverse complaint-video datasets, stronger grounding controls for retrieval, and evaluation procedures that reduce dependence on heuristic sentiment tools and external LLM judges.