Recently, there has been a growing trend of consumers seeking product information from video platforms such as YouTube. However, when viewing multiple review videos about the same product, viewers often encounter redundant information, resulting in wasted time. To address these issues, we use BERTopic to eliminate repetitive video content and address the problem of missing subtopics, which has been a limitation of traditional video summarization methods. Subsequently, the topic-aware video contents are summarized using the BART model. The ROUGE metric was used to evaluate the model proposed in this paper, and the experimental results showed improved results compared to previous research.
1. Introduction
The increase in the use of the YouTube platform as a main source of information mentioned earlier is the same for product reviews. Product reviews are a way for buyers to learn about the pros and cons of a product and its features before making a purchase. If a product review is long, it takes a lot of time to read through the content of the entire review. For this reason, there is a need for product review summaries that minimize the length of product reviews while preserving their content. [12], [13] With the increase of product review videos on the YouTube platform, there has been an increase in the number of multiple product review videos for a single product. Multi-video product review summaries provide product reviews from different perspectives for buyers who want to purchase a product and help sellers or companies to improve their quality of products and services [14].
In this paper, we aim to identify the semantic meanings and relationships of product review videos through subtitle summaries and provide product reviews from different perspectives through multi-video product review summaries.
Extractive summaries are summaries that attempt to extract the entire summarized document into sentences, phrases, and words from the source document. Generative summaries are summaries in which the summarized document generates a summary using words, sentences, and phrases that are not present in the original document. [15]In this paper, we use extractive summarization to preserve the review video creator's intent by using sentences extracted from the original document to increase the reliability of the review.
Topic modeling is a statistical model that identifies the topics of a set of documents in the field of machine learning and natural language processing. It analyzes large amounts of textual data to help identify the embedded semantic structure of the text and summarize its key content.[16]In this paper, we use topic modeling to identify various embedded topics and subtopics.
Ansamma et al. (2017) [17]maximized relevance and minimized redundancy by representing them as word vectors, and multi-document extractive summarization using Latent Semantic Analysis (LSA) and Non-Negative Matrix Factorization (NMF) as objective functions. Alrumiah et al. (2022) [18]is a single summary of lecture video subtitles using LDA. By generating a keyword list using LDA and extracting sentences from the original document that contain at least one word from the keyword list, they improved the summarization performance in terms of length and quality compared to previous studies. Miller et al. (2019) [19]summarized a single lecture video using BERT. This is the first study in which a large language model (LLM) is used for video summarization, and the LLM model, BERT, is used to embed the input sentences. Then, the n most centroid sentences were selected using k-means to summarize the lecture video subtitles. The performance evaluation was compared with TextRank, but since there was not Golden summary, it was not evaluated using evaluation metrics.
In the data preprocessing stage, we remove emoticons, onomatopoeia, and onomatopoeia, which are unnecessary information in this summary. For onomatopoeia and onomatopoeia containing more than one consonant and vowel, we removed all but one word. If there were missing values and subtitles other than Korean, the columns were deleted. In addition, we removed stop words, which are information that does not add meaningful information to the sentence. We used the list of Korean stop words provided by NLTK.
BERTopic is a popular BERT-based algorithm for topic modeling. It utilizes Transformer and cTF-IDF to generate dense clusters while maintaining important words in the topic. The process of BERTopic includes three steps: Document Embedding, Document Clustering, and Topic representation. Each step is designed to be independent and can be customized according to the purpose [20].
In this paper, when comparing topic modeling algorithms, we found that BERTopic shows equally good results in Topic Coherence and Topic Diversity. This means that it generates topics with diversity while maintaining the coherence of the document's topics. In addition, BERTopic has the advantage of requiring relatively little time even as the size of the vocabulary increases. For these reasons, we chose BERTopic for this paper.
BART is a denoising autoencoder based on the Transformer architecture.[21]The corrupted text is input to the encoder and the text representation learned by the encoder is sent to the decoder, which recovers the original undamaged text. BART is a model that can be trained by combining natural language understanding and generation, and performs well on summarization tasks.
This is the process of converting summarized text to video. The process extracts the number of the video where the text is located and the timeline of the video where the text is located. The video is then cut to include the summarized content. The process is iterated over the number of summarized texts, and then the summarized videos are joined together into a single video. Figure 3 is an image that captures the summarized video created through this process.
For Korean data, there is no existing product review data for multi-video summarization, thus we created a dataset in this paper. The dataset is composed of 271 Korean product review videos of 22 different products such as cell phones, game consoles, and robot vacuum cleaners. Shown in Figure 4 below is the shape of the dataset. Each column of the data is in the format of product name, video title, video link, video subtitle, time of video matching with subtitle, and answer data. The answer data was generated randomly.
For evaluation metrics, we used Topic Coherence and Topic Diversity to evaluate Topic modeling, and Rouge for summary evaluation.
Topic Coherence evaluates Normalized Pointwise Mutual Information (NPMI) [22], a method proposed by Röder et al. (2015) [23]. The evaluated result has a value between -1 and 1. The closer it is to 1, the more the coherence of the topic.
⃗ ( ′) = { ∑
( , ) } ∈ ′ =1,…,| |
Let ′be a vector representing the similarity between words in the corpus. ⃗ ( ′) computes the similarity of each word in ′ to the other words and represents it as a vector.
NPMI is a metric that measures the relevance of two words , . ( , ) is the probability that two words , co-occurence. ( ) is the probability of occurrence and ( ) is the probability of occurrence. represents an exponent and serves to weight the NPMI values exponentially.
Φ (⃗⃗ , ⃗⃗⃗ ) =
| | ∑
=1 ∙ ‖⃗⃗ ‖2 ∙ ‖⃗⃗⃗ ‖2 represents the dimension of the vector. ‖⃗⃗ ‖ 2 represents the L2 norm of vector ⃗⃗ , ‖⃗⃗⃗ ‖2 ⃗⃗ , ⃗⃗⃗ by dividing the dot product of the vectors by the L2 norm of the vectors. represents the L2 norm of vector ⃗⃗⃗ , and Φ (⃗⃗ , ⃗⃗⃗ ) computes the similarity between two vectors
Dieng et al. (2020) [24] proposed a metric to measure the diversity of vocabulary or words in the topics of a certain topic model. The metric is the proportion of unique words, excluding duplicate words, that exist in the topics for all topics, and the measure ranges from 0 to 1. 0 represents duplicate topics and 1 represents various topics.
|{
}| is the size of the union of unique words in all topics. topk is the selected number of top topics, and topics is the number of total topics. = |{ ∙ | }| |
( , ) = ROUGE − N = ∑ ∈
∑ − ∑ ∈ ∑ − ∈ ∈
ℎ( − ( − )
)
(
ROUGE (Recall-Oriented Understudy for Gisting Evaluation) (Lin, 2004) is a co-occurrence
statistical measure of N-grams and is to be defined as follows (
) is the number of N-grams included in the reference summary. These ROUGE
metrics can generate three scores: Recall, Precision, and F-measure [25].
ℎ(
−
) is the
(
We calculated the TC and TD of three models, the BERTopic model with distilues-basemultilngual-cased-v1 as the embedding model, the Combined Topic Models (CTM) model, and the LDA model, and found that BERTopic shows the best performance, as shown in Table 2. This shows that the BERTopic model generates various topics and coherent topics in one topic. We calculated the TC and TD of three models, the BERTopic model with distilues-basemultilngual-cased-v1 as the embedding model, the Combined Topic Models (CTM) model, and the LDA model, and found that BERTopic shows the best performance, as shown in Table 2. This shows that the BERTopic model generates various topics and coherent topics in one topic. 4.4. Bart Summarization step uses BART to generate a summary for each topic generated by BERTopic. After sorting the sentences with the highest cosine similarity, we extracted the top n sentences to summarize each topic. In this paper, we set n to 3 arbitrarily. The performance comparison was performed with TextRank [26] and Bert-extractive-summarizer [19]. The results can be seen in Table 3. In precision, TextRank and Bert-extractive-summarizer showed good performance. But, in recall and f1-measure, we can show that the proposed model has the best performance. This means that the proposed model performs the most equally well compared to the other models.
In this paper, we reduce the duplication of data by grouping the same or similar information into one topic through Topic Modeling. In addition, we compared models such as CTM and LDA, and selected BERTopic with the best topic diversity and topic cohesion to generate coherent and diverse topics. This process reduces wasted time searching for information, provides users with multiple perspectives on the product, and helps eliminate information loss in long videos. We used BART for summarization. In comparison to the previous research such as Text Rank and Bert-extractive-summarizer, ROUGE showed the best performance in recall and F1 measures.
This research was supported by Basic Science Research Program through the National Research Foundation of Korea (NRF), funded by the Ministry of Education, Science and Technology (NRF2021R1G1A1006381). [17] John, Ansamma, P. S. Premjith, and M. Wilscy. "Extractive multi-document summarization using population-based multicriteria optimization." Expert Systems with Applications 86 (2017): 385-397. [18] S. S. Alrumiah and A. A. Al-Shargabi, "Educational videos subtitles’ summarization using latent dirichlet allocation and length enhancement," Computers, Materials & Continua, vol. 70, no.3, pp. 6205–6221, 2022. [19] Miller, Derek. "Leveraging BERT for extractive text summarization on lectures." arXiv preprint arXiv:1906.04165 (2019). [20] Grootendorst, Maarten. "BERTopic: Neural topic modeling with a class-based TF-IDF procedure." arXiv preprint arXiv:2203.05794 (2022). [21] Lewis, Mike, et al. "Bart: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension." arXiv preprint arXiv:1910.13461 (2019). [22] Bouma, Gerlof. "Normalized (pointwise) mutual information in collocation extraction." Proceedings of
GSCL 30 (2009): 31-40. [23] Röder, Michael, Andreas Both, and Alexander Hinneburg. "Exploring the space of topic coherence measures." Proceedings of the eighth ACM international conference on Web search and data mining. 2015. [24] Dieng, Adji B., Francisco JR Ruiz, and David M. Blei. "Topic modeling in embedding spaces." Transactions of the Association for Computational Linguistics 8 (2020): 439-453. [25] Lin, Chin-Yew. "Rouge: A package for automatic evaluation of summaries." Text summarization branches out. 2004. [26] Mihalcea, Rada, and Paul Tarau. "Textrank: Bringing order into text." Proceedings of the 2004 conference on empirical methods in natural language processing. 2004