---
title: 'ComVID: Multimodal Complaint Benchmark'
url: https://www.emergentmind.com/topics/comvid
type: topic
---

# ComVID: Multimodal Complaint Benchmark

ComVID is a multimodal video complaint dataset introduced together with Complaint Description from Videos (CoD-V), a generation task that formalizes complaint-focused text generation from video evidence, textual prompting, and emotion labels. In the 2025 formulation, ComVID contains 1,175 complaint videos with gold complaint descriptions and emotion annotations, and it is intended to support systems that do more than summarize or caption video: they must preserve the “complaint nature,” ground claims in visible defect evidence, and reflect the complainer’s affect [2509.19952]. The same label also appears in a separate chest-CT COVID-19 identification context, where “ComVID” is used for slice-level COVID-19 identification from CT images; the two usages are unrelated in task, modality, and evaluation [2207.04212].

## 1. CoD-V as a complaint-generation task

CoD-V is defined as generating a coherent, complaint-focused narrative $Y$ from a video $V$, a textual prompt $P$, and an emotion label $e$, i.e., $Y = F(V, P, e)$ [2509.19952]. The task is positioned within complaint mining rather than generic video-language modeling. Its objective is not merely to describe visible events or compress video content, but to express a user complaint about product defects and service issues as evidenced in the video, in a tone aligned with the user’s emotional state.

The task’s novelty is specified along three axes. First, it grounds textual complaints in fine-grained visual cues, such as a broken component. Second, it retains complaint-specific aspects including quality, functionality, defective components, delays, refunds, and missing items. Third, it conditions generation on emotion labels so that the resulting narrative reflects dissatisfaction, blame, frustration, or disappointment. This directly differentiates CoD-V from standard video captioning and video summarization, which can produce semantically plausible but complaint-insensitive outputs that omit defect specificity or emotional stance [2509.19952].

This framing has methodological consequences. Standard captioning metrics can reward fluent descriptions that fail to encode defect semantics, while complaint mining requires preservation of aspect-specific and affective information. A plausible implication is that CoD-V should be understood as a task at the intersection of multimodal grounding, affect-aware NLG, and application-oriented complaint mining rather than as a straightforward extension of captioning.

## 2. Dataset construction, schema, and annotation

ComVID is built from public Amazon review videos, scraped with BeautifulSoup from 1- and 2-star reviews. The stored metadata per sample include review ID, rating, review text, aspects, domain name, product name, and m3u8 video links; the videos were transcoded to mp4 for accessibility [2509.19952]. After selection and curation, the final dataset comprises 1,175 videos, each paired with a gold complaint description and an emotion label.

| Domain | Videos |
|---|---:|
| Electronics | 655 |
| Household | 273 |
| Fashion | 202 |
| Others | 45 |

The dataset targets four domains—Fashion, Electronics, Household, and Others—and organizes complaint content using a cross-domain taxonomy of aspects: Quality, Functionality, Defective, Design, Missing, Refund, Performance, and Delay. The language distribution is primarily English, with six samples including non-English (Hindi). Each item carries an emotion label drawn from $\{ \text{dissatisfaction}, \text{blame}, \text{frustration}, \text{disappointment} \}$ [2509.19952].

Annotation proceeded in two phases. Phase-1 involved five expert linguists: one PhD-level annotator (Category A), two Masters-level annotators (Category B), and two undergraduate annotators (Category C), all bilingual in Hindi and English. Category A created 50 gold-standard references. Category B and C generated descriptions for half the dataset each, and the remaining two annotators reviewed and cross-validated all samples for factual alignment with video content, coherence, and guideline adherence. Only samples with acceptance tags became gold descriptions. Phase-2 involved independent emotion labeling by Category A and B annotators, followed by consensus resolution. Annotators were compensated at $0.50$ per sample. Emotion labeling achieved Fleiss’ $\kappa = 0.64$, reported as substantial agreement; Cohen’s $\kappa$ is not reported [2509.19952].

The paper also states explicit ethical and licensing conditions. ComVID is derived from publicly accessible Amazon review content, avoids collecting personally sensitive details beyond what users already published, and is released under a Creative Commons license, CC BY-NC-ND, which restricts commercial reuse and derivative modifications. The authors note residual bias sources, especially domain skew toward electronics and English predominance, while arguing that multi-annotator consensus and domain balancing mitigate some of these issues [2509.19952].

## 3. Evaluation and the Complaint Retention metric

A central contribution associated with ComVID is Complaint Retention (CR), a metric proposed to distinguish CoD-V from generic summary generation and video description [2509.19952]. CR aggregates three sub-scores: normalized sentiment, emotion intensity, and aspect retention.

The normalized VADER sentiment for sentence $i$ is defined as
$$
S_{\text{N VADER}, i} = \frac{S_{\text{VADER}, i} + 1}{2},
$$
and the average per-sample sentiment over $N_k$ sentences is
$$
S_{\text{N VADER}, k} = \frac{1}{N_k} \sum_{i=1}^{N_k} S_{\text{N VADER}, i}.
$$

Emotion Score uses Text2Emotion with the categories Happy $(H)$, Angry $(A)$, Surprise $(S)$, and Fear $(F)$:
$$
ES_k = \frac{H_k + A_k + S_k + F_k}{4}.
$$

Aspect Score measures the fraction of ground-truth aspects present in the predicted text $T_p$, using a GPT-4 “Yes/No” classifier:
$$
AS_k = \frac{\sum_{i=1}^{N_k} \mathbf{1}(\text{Aspect}_i \in T_p)}{N_k}.
$$

The final CR score averages these three components. The paper’s stated intuition is that a well-formed complaint should exhibit negative sentiment intensity aligned with dissatisfaction, encode emotional tone, and preserve complaint aspects such as “scroll wheel not working” or “right earcup broken” [2509.19952]. A generic summary with neutral sentiment or omitted aspects should therefore receive a lower CR score.

CR is evaluated alongside standard NLG metrics: BLEU, ROUGE, METEOR, BERTScore, MoverScore, Perplexity, Flesch Reading Ease, and the Coleman–Liau Index. The paper emphasizes that only CR and VADER scores clearly discriminate complaint-specific generation from generic summaries or descriptions, because lexical and semantic overlap metrics alone do not directly penalize emotional mismatch or aspect omission [2509.19952].

## 4. Modeling pipeline and retrieval-augmented generation

The ComVID study evaluates several video-language baselines and state-of-the-art models, including LSTM+VGG16, LSTM+ResNet50, BLIP-VQA-BASE, Qwen2-VL-7b, Gemma3-12b, VideoLLaVA2-7b, and VideoLLaMA2-7b [2509.19952]. Its principal modeling contribution is a multimodal Retrieval-Augmented Generation (RAG) embedded VideoLLaMA2-7b model.

The system is described in two steps. Step 1 is Multimodal Retrieval (MR). The retrieval corpus is a large-scale Amazon product review dataset containing 75.26M reviews, with the constraint that each user and product has at least five reviews. This corpus is filtered to genuine complaints using three criteria: rating = 1 star, text length $> 150$ characters, and at least one image. CLIP encoders are used for text and images. The text embedding is defined as $e_t = \text{TextEncoder}(T)$, where $T$ is title plus main text; image embeddings are averaged to form $\bar{e_i}$; and the multimodal review embedding is the average of text and image representations, indexed with FAISS for similarity search. For the video query, four keyframes are extracted using GMFlow and encoded with CLIP’s image encoder, then averaged to obtain $q_{\text{video}}$. Product aspect text is also encoded via CLIP to obtain $e_{\text{textual}}$. These are fused into a final query embedding used to retrieve top-$k$ nearest multimodal neighbors, supplying complaint-grounded context aligned to the product and aspect [2509.19952].

Step 2 is supervised fine-tuning and generation. VideoLLaMA2-7b is fine-tuned on supervised video-text pairs. During inference, the prompt is enriched with the user’s emotion label and the top-$k$ retrieved complaints from MR so that generation is grounded in both defect semantics and tone. The reported training details are LoRA rank = 64, LoRA alpha = 128, batch size = 16, learning rate $= 2 \times 10^{-4}$ with cosine scheduling, multimodal projector learning rate $= 2 \times 10^{-5}$, and 10 epochs [2509.19952].

This architecture makes ComVID more than a static dataset. It also functions as a benchmark for studying how retrieval and affect conditioning interact in multimodal generation. The design suggests that complaint generation benefits from exemplar-based contextualization in addition to direct video understanding.

## 5. Empirical findings, ablations, and failure modes

The main quantitative results indicate that fine-tuned VideoLLaMA2-7b+MR outperforms the reported baselines across ROUGE, BLEU, BERTScore, METEOR, and readability metrics [2509.19952]. With emotion conditioning, the proposed model achieves ROUGE-1/2/L = 0.59/0.47/0.56, BLEU-1/2 = 0.69/0.63, BLEU overall = 0.59, BERTScore = 0.93, Flesch Reading Ease $\approx 79.58$, and lower perplexity than many baselines, reported as PS $\approx 97.16$.

The ablation results attribute gains to both emotion conditioning and retrieval. Emotion injection significantly improves metrics and the perceived quality of complaint narratives. Integrating MR yields 3–4% improvements over fine-tuned VLMs without RAG, which the paper interprets as improved factual grounding and aspect retention. In zero-shot comparisons across summary generation, video description, and CoD-V, the proposed model attains higher CR, with a reported value of 0.62, indicating better preservation of complaint sentiment and aspects than generic generation tasks [2509.19952].

Qualitative analysis reinforces these quantitative findings. In a representative mouse-defect example, the proposed model explicitly states that “the mouse is defective,” that a damaged product was received, and that the “scrolling button” is not working properly, closely matching the gold complaint. By contrast, BLIP-VQA-BASE is described as more generic, and VideoLLaVA2-7b is reported to misattribute the issue by discussing product age rather than defect specifics [2509.19952].

The error analysis, however, identifies nontrivial failure modes. The proposed system can hallucinate ancillary claims, such as blaming a retailer or mentioning customer care difficulty, even when such content is not evidenced in the video. The paper also notes that retrieval can introduce context leakage or overgeneralization, and that emotion conditioning may amplify negative sentiment beyond the available evidence if not carefully controlled. Human evaluation on 80 randomly sampled test instances reports higher Information Preservation Ratings for Adequacy, Emotion Encoding, and Visual Content Alignment than BLIP-VQA-BASE and VideoLLaVA2-7b, but the train/validation/test splits are not explicitly enumerated in the paper [2509.19952].

## 6. Applications, limitations, and terminological overlap

The immediate applications described for ComVID are customer service platforms and e-commerce tooling. Systems trained on the dataset could auto-generate complaint text from user-uploaded videos, accelerate triage and resolution, and assist users with limited literacy or difficulty expressing complaints in text [2509.19952]. The broader claim is not that video description is sufficient for complaint mining, but that complaint-aware generation requires explicit preservation of defect aspects and affective stance.

The study’s limitations are also explicit. ComVID remains relatively small at 1,175 videos, is skewed toward electronics, has limited language diversity, and uses a discrete emotion taxonomy consisting of dissatisfaction, blame, frustration, and disappointment. The authors suggest enlarging and diversifying the dataset across languages and domains, modeling finer-grained emotion taxonomies and intensity, strengthening grounding constraints to curb hallucinations, and extending the framework to adjacent domains such as healthcare incident reporting, content moderation, and real-time applications [2509.19952].

A separate use of the term “ComVID” appears in a 2022 medical-imaging study on COVID-19 disease identification from chest CT images. There, the term refers to slice-level CT-based COVID-19 identification using a custom CNN and VGG16 on a combined dataset of 14,320 CT images, with reported accuracies of 96.34% for CNN and 96.99% for VGG16 in the Results section, alongside a minor inconsistency in the Conclusion, which reports 97.40% for VGG16 [2207.04212]. That usage concerns radiological classification, not complaint mining. The overlap is terminological rather than conceptual: one ComVID denotes a multimodal video complaint benchmark, while the other denotes COVID-19 identification on chest CT images.

Source: https://www.emergentmind.com/topics/comvid