| { |
| "paper_id": "Y18-1007", |
| "header": { |
| "generated_with": "S2ORC 1.0.0", |
| "date_generated": "2023-01-19T13:36:12.443624Z" |
| }, |
| "title": "Predicting the Genre and Rating of a Movie Based on its Synopsis", |
| "authors": [ |
| { |
| "first": "Varshit", |
| "middle": [], |
| "last": "Battu", |
| "suffix": "", |
| "affiliation": {}, |
| "email": "battu.varshit@research.iiit.ac.in" |
| }, |
| { |
| "first": "Vishal", |
| "middle": [], |
| "last": "Batchu", |
| "suffix": "", |
| "affiliation": {}, |
| "email": "vishal.batchu@students.iiit.ac.in" |
| }, |
| { |
| "first": "Rohit", |
| "middle": [], |
| "last": "Rama", |
| "suffix": "", |
| "affiliation": {}, |
| "email": "" |
| }, |
| { |
| "first": "Krishna", |
| "middle": [], |
| "last": "Reddy", |
| "suffix": "", |
| "affiliation": {}, |
| "email": "" |
| }, |
| { |
| "first": "Radhika", |
| "middle": [], |
| "last": "Reddy", |
| "suffix": "", |
| "affiliation": {}, |
| "email": "radhika.mamidi@iiit.ac.in" |
| }, |
| { |
| "first": "", |
| "middle": [], |
| "last": "Mamidi", |
| "suffix": "", |
| "affiliation": {}, |
| "email": "" |
| } |
| ], |
| "year": "", |
| "venue": null, |
| "identifiers": {}, |
| "abstract": "Movies are one of the most prominent means of entertainment. The widespread use of the Internet in recent times has led to large volumes of data related to movies being generated and shared online. People often prefer to express their views online in English as compared to other local languages. This leaves us with a very little amount of data in languages apart from English to work on. To overcome this, we created the Multi-Language Movie Review Dataset (MLMRD). The dataset consists of genre, rating, and synopsis of a movie across multiple languages, namely Hindi, Telugu, Tamil, Malayalam, Korean, French, and Japanese. The genre of a movie can be identified by its synopsis. Though the rating of a movie may depend on multiple factors like the performance of actors, screenplay, direction etc but in most of the cases, synopsis plays a crucial role in the movie rating. In this work, we provide various model architectures that can be used to predict the genre and the rating of a movie across various languages present in our dataset based on the synopsis.", |
| "pdf_parse": { |
| "paper_id": "Y18-1007", |
| "_pdf_hash": "", |
| "abstract": [ |
| { |
| "text": "Movies are one of the most prominent means of entertainment. The widespread use of the Internet in recent times has led to large volumes of data related to movies being generated and shared online. People often prefer to express their views online in English as compared to other local languages. This leaves us with a very little amount of data in languages apart from English to work on. To overcome this, we created the Multi-Language Movie Review Dataset (MLMRD). The dataset consists of genre, rating, and synopsis of a movie across multiple languages, namely Hindi, Telugu, Tamil, Malayalam, Korean, French, and Japanese. The genre of a movie can be identified by its synopsis. Though the rating of a movie may depend on multiple factors like the performance of actors, screenplay, direction etc but in most of the cases, synopsis plays a crucial role in the movie rating. In this work, we provide various model architectures that can be used to predict the genre and the rating of a movie across various languages present in our dataset based on the synopsis.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Abstract", |
| "sec_num": null |
| } |
| ], |
| "body_text": [ |
| { |
| "text": "As the amount of data present online increases exponentially day by day, we have reached a point where a human cannot comprehend all of it in a meaningful manner due to its sheer size. This lead to work on automated recommender systems. The main issue with these kinds of methods is that not all the information is present online and all the information present need not be correct. Automated movie genre and rating prediction have a lot of applications. We can recommend same genre movies based on his previous watch history. Genre of a movie can be identified by its synopsis. Recommending a movie only based on its genre is not a good idea as the same genre can have both good and bad movies. So Recommending movies based on both genre and rating would result in a proper recommendation system. But the main problem here is that people do not often tend to rate the movie they watch, thus automated rating prediction would be of great help for recommendation systems. Though the rating of a movie depends on multiple factors like actors, screenplay, direction etc. but that information is very difficult to capture through available data. In most of the cases, synopsis of the movie plays a crucial impact on audience rating. In this paper, we propose multiple deep-learning based methods to predict the genre and rating of a movie based on its synopsis.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Introduction", |
| "sec_num": "1" |
| }, |
| { |
| "text": "There is a very little amount of data in languages apart from English to work on. To overcome this, we created the Multi-Language Movie Review Dataset (MLMRD). The dataset consists of genre, rating, and synopsis of a movie across multiple languages, namely Hindi, Telugu, Tamil, Malayalam, Korean, French, and Japanese. Balance in the dataset is not that good because nowadays movies in specific languages tend to belong to only specific genres due to various reasons like movie collections, ease of making etc. For example, no documentary movies are present in Telugu as such movies make fewer collections at Tollywood box office. Tamil Malayalam French Japanese Korean Genre Action 230 45 21 56 1,314 763 15 Comedy 60 35 27 25 2,602 15 6 Crime 8 10 15 0 0 0 0 Drama 47 88 62 43 3,425 2,798 60 Family 18 0 0 21 0 763 0 Horror 20 0 17 9 208 278 0 Romance 133 42 19 14 127 0 2 Thriller 43 33 20 18 532 0 14 Documentary 0 0 0 0 833 38 19 Rating 1 10 16 7 19 766 267 1 2 140 41 83 ", |
| "cite_spans": [], |
| "ref_spans": [ |
| { |
| "start": 632, |
| "end": 1067, |
| "text": "Tamil Malayalam French Japanese Korean Genre Action 230 45 21 56 1,314 763 15 Comedy 60 35 27 25 2,602 15 6 Crime 8 10 15 0 0 0 0 Drama 47 88 62 43 3,425 2,798 60 Family 18 0 0 21 0 763 0 Horror 20 0 17 9 208 278 0 Romance 133 42 19 14 127 0 2 Thriller 43 33 20 18 532 0 14 Documentary 0 0 0 0 833 38 19 Rating 1 10 16 7 19 766 267 1 2 140 41 83", |
| "ref_id": "TABREF1" |
| } |
| ], |
| "eq_spans": [], |
| "section": "Introduction", |
| "sec_num": "1" |
| }, |
| { |
| "text": "Work has been done in related areas in the past. Basu et al. (Basu et al., 1998) propose an inductive learning approach to predict user preferences. Huang et al. (Huang and Wang, 2012) propose a movie genre classification system using a meta-heuristic optimization algorithm called Self-Adaptive Harmony Search. Rasheed et al. (Rasheed and Shah, 2002) present a method to classify movies on the basis of audio-visual cues present in previews which contain important information about the movie. Zhou et al. (Zhou et al., 2010) present a method for movie genre categorization of movie trailers, based on scene categorization. Gabriel S. Simoes et al. (Sim\u00f5es et al., 2016) explored CNNs in the context of movie trailers genre classification. Firstly, a novel movie trailers dataset with more than 3500 trailers was publicly released. Secondly, a novel classification method was done which encapsulates a CNN architecture to perform movie trailer genre classification, namely CNN-MoTion. Chin-Chia Michael Yeh et al. (Yeh and Yang, 2012) concerns the development of a music codebook for summarizing local feature descriptors computed over time. With the new supervised dictionary learning algorithm and the optimal settings inferred from the performance study, they achieved the state-of-the-art accuracy of music genre classification. Aida Austin et al. (Austin et al., 2010) 6 for Japanese and www.allocine.fr 7 for French. We scraped the rating, genre, and synopsis of every movie from each website. Due to lack of resources and not much data is available in the specific language script, we could only mine a small amount of data. There aren't a lot of regional sites available that have trustworthy information to collect data from. Hence a lot of the languages have only a small number of data points in our dataset. However, we have ensured that the data collected, although small, is valid and collected from reputed movie review sites. We believe that having a small but strong and correct dataset is better than having a large dataset with a lot of noise and hence did not include other sites that did not have much reputation. The anonymized code can be found at https://goo.gl/nbWD9s and the dataset can be found at https://goo.gl/xpFv9q.", |
| "cite_spans": [ |
| { |
| "start": 49, |
| "end": 80, |
| "text": "Basu et al. (Basu et al., 1998)", |
| "ref_id": null |
| }, |
| { |
| "start": 149, |
| "end": 184, |
| "text": "Huang et al. (Huang and Wang, 2012)", |
| "ref_id": "BIBREF3" |
| }, |
| { |
| "start": 312, |
| "end": 351, |
| "text": "Rasheed et al. (Rasheed and Shah, 2002)", |
| "ref_id": "BIBREF6" |
| }, |
| { |
| "start": 507, |
| "end": 526, |
| "text": "(Zhou et al., 2010)", |
| "ref_id": null |
| }, |
| { |
| "start": 625, |
| "end": 671, |
| "text": "Gabriel S. Simoes et al. (Sim\u00f5es et al., 2016)", |
| "ref_id": null |
| }, |
| { |
| "start": 1015, |
| "end": 1035, |
| "text": "(Yeh and Yang, 2012)", |
| "ref_id": "BIBREF7" |
| }, |
| { |
| "start": 1353, |
| "end": 1374, |
| "text": "(Austin et al., 2010)", |
| "ref_id": "BIBREF0" |
| }, |
| { |
| "start": 1375, |
| "end": 1376, |
| "text": "6", |
| "ref_id": null |
| } |
| ], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Related Work", |
| "sec_num": "2" |
| }, |
| { |
| "text": "The websites mentioned above have links to the synopsis of each movie along with the genre and rating in that web page. We first saved those links and then used beautiful soup to scrape the web page and get the synopsis, genre, and rating of the movie. ", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Data Extraction", |
| "sec_num": "3.1" |
| }, |
| { |
| "text": "After collecting all the required data, we had to preprocess the data to cluster genres into classes. Since the data collected had different classes in each language we merged similar classes into one broader class as explained in Section 3.2.1. Finally, we were left with 9 classes viz. action, comedy, crime, drama, family, horror, romance, thriller and documentary for each language. The details are mentioned in Table 1 . We only added movies having all the three -genre, rating and synopsis into the dataset and ignored movies which were missing information. To validate the data, we performed a manual inspection at various data points selected at random to ensure the ratings and genres are valid and not erroneous. Here a synopsis, its respective rating and it's genre is referred as a data point. Once this was done we shuffled the data before passing it through the model. Shuffling the data helps make the training and test sets more representative of the overall distribution of the dataset. We then split the data into two parts containing 80% and 20% of the entire data for training and testing respectively.", |
| "cite_spans": [], |
| "ref_spans": [ |
| { |
| "start": 416, |
| "end": 423, |
| "text": "Table 1", |
| "ref_id": "TABREF1" |
| } |
| ], |
| "eq_spans": [], |
| "section": "Preprocessing", |
| "sec_num": "3.2" |
| }, |
| { |
| "text": "The original data had several kinds of genres. We grouped all relevant genres together to finally end up with 9 different classes of genres as mentioned in Table 2. For example Autobiography, Biopic etc were put into the Documentary class. Romantic-Comedy as the name suggests could be part of the Romance group but when we manually inspected a few data points at random they were more suited for the comedy genre and hence put it there.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Grouping of genres", |
| "sec_num": "3.2.1" |
| }, |
| { |
| "text": "There are 14,991 entries in the dataset we compiled. The language-wise distribution of entries and words per language are mentioned in Table 3 . This is a big dataset covering a total of seven languages belonging to different language families. Each language has different average lengths of the synopsis in terms of the number of words. For example Hindi has 253 entries and 1,67,842 words whereas Tamil has 181 entries and 1,01,904 words. ", |
| "cite_spans": [], |
| "ref_spans": [ |
| { |
| "start": 135, |
| "end": 142, |
| "text": "Table 3", |
| "ref_id": "TABREF7" |
| } |
| ], |
| "eq_spans": [], |
| "section": "Statistics", |
| "sec_num": "3.3" |
| }, |
| { |
| "text": "We predict the genre and rating of a movie based on its synopsis alone. Genre prediction deals with 9 output classes as shown in Table 2 . We treat rating prediction as a classification task rather than regression. We round the ratings leaving us with 5 classes that we try to predict.", |
| "cite_spans": [], |
| "ref_spans": [ |
| { |
| "start": 129, |
| "end": 136, |
| "text": "Table 2", |
| "ref_id": "TABREF5" |
| } |
| ], |
| "eq_spans": [], |
| "section": "Genre and Rating Prediction", |
| "sec_num": "4" |
| }, |
| { |
| "text": "Each character of the input synopsis is converted to a vector dynamically using an Embedding layer at the inputs to the networks. These character vectors are then passed along to various convolution and recurrent networks. Using a one-hot encoded representation of the characters also gave similar accuracies.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Character Embeddings", |
| "sec_num": "4.1" |
| }, |
| { |
| "text": "Convolution networks: The input to the CNN (LeCun et al., 2015) consists of all the character embeddings stacked as filters which are then passed along the network to predict an output genre/rating class. The network as mentioned in Figure 1 has a branched structure where filters of various sizes are used in the convolution layers in each of the branches and the outputs are concatenated before being passed onto fully connected layers to predict the output genre/rating class. Figure 2, we feed in character vectors one at a time as input and the predicted output is passed forward to multiple fully connected dense layers which predict the output genre/rating class.", |
| "cite_spans": [], |
| "ref_spans": [ |
| { |
| "start": 233, |
| "end": 241, |
| "text": "Figure 1", |
| "ref_id": "FIGREF0" |
| }, |
| { |
| "start": 480, |
| "end": 486, |
| "text": "Figure", |
| "ref_id": null |
| } |
| ], |
| "eq_spans": [], |
| "section": "Character Embeddings", |
| "sec_num": "4.1" |
| }, |
| { |
| "text": "Each word in the input is converted into a vector. These vectors are generated either dynamically using an Embedding layer or statically using Gensim (Rehurek and Sojka, 2010). These generated vectors are used as inputs to convolution and recurrent networks similar to how character encodings were used.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Word Embeddings", |
| "sec_num": "4.2" |
| }, |
| { |
| "text": "Sentence vectors were generated using Doc2Vec (Le and Mikolov, 2014) . Doc2Vec takes all the sentences at once and generates sentence vectors for them. However, this requires all the data to be fed into Doc2Vec i.e both train and test sentences and hence this cannot be performed on unseen data.", |
| "cite_spans": [ |
| { |
| "start": 46, |
| "end": 68, |
| "text": "(Le and Mikolov, 2014)", |
| "ref_id": "BIBREF4" |
| } |
| ], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Sentence Embeddings", |
| "sec_num": "4.3" |
| }, |
| { |
| "text": "Fully connected networks: Since the entire synopsis is encoded using a single vector, we pass the vector through a fully connected network which predicts the output genre/rating class and convolution/recurrent networks provide no benefits here.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Sentence Embeddings", |
| "sec_num": "4.3" |
| }, |
| { |
| "text": "We observed that different types of embeddings performed well for different languages, for example, word embeddings for Telugu and Hindi, sentence embeddings for Tamil etc. Hence, concatenating all the three embeddings namely character, word and sentence embeddings and pass them through different models so that there can be an increase in the accuracy as the network chooses important parts of these embeddings.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Concatenated Embeddings", |
| "sec_num": "4.4" |
| }, |
| { |
| "text": "We performed numerous experiments using various deep learning models including convolution and recurrent based networks with character, word, and sentence level embeddings for inputs. We also compare our proposed models with some of the popular traditional approaches such as SVMs (Cortes and Vapnik, 1995) and Random Forests (Svetnik et al., 2003) and show that deep learning based methods beat them by large margins as shown in Tables 4 and 5.", |
| "cite_spans": [ |
| { |
| "start": 281, |
| "end": 306, |
| "text": "(Cortes and Vapnik, 1995)", |
| "ref_id": null |
| }, |
| { |
| "start": 326, |
| "end": 348, |
| "text": "(Svetnik et al., 2003)", |
| "ref_id": "BIBREF6" |
| } |
| ], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Experiments", |
| "sec_num": "5" |
| }, |
| { |
| "text": "We use Keras with the Tensorflow backend to perform all our experiments. We use a GeForce GTX-1080Ti GPU in order to train our models (Each model takes less than 15 minutes to complete training). We use dropouts at various locations in the networks to reduce over-fitting. Categorical cross entropy loss is used as the loss function along with the Adam optimizer for training all the networks. We observe that dynamic embeddings perform better than static embeddings in all word based models and hence we use embedding layers in all the models instead of using Gensim or GloVe word vectors. ReLU activations are used throughout the networks except for the last layers which use SoftMax activations in all the models. The code provided along with the paper has further implementation details. Figure 3 : Hybrid model based on stacking three models to predict genre/rating with word, character and sentence embeddings as the input considers synopsis to perform the prediction. We used Dropouts at various places for regularization.", |
| "cite_spans": [], |
| "ref_spans": [ |
| { |
| "start": 792, |
| "end": 800, |
| "text": "Figure 3", |
| "ref_id": null |
| } |
| ], |
| "eq_spans": [], |
| "section": "Experimental Details", |
| "sec_num": "5.1" |
| }, |
| { |
| "text": "SVMs are commonly used for recognition and regression analysis of data. Considering features from the reviews as inputs, they try to classify them into one of the genre and rating classes. We run a trained Doc2Vec model using the entire review as an input and that provides us a 300-dimensional embedding that we use as an input to the SVM.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "SVM", |
| "sec_num": "5.1.1" |
| }, |
| { |
| "text": "Similar to the features we use for the SVM, we use the Doc2Vec embeddings of reviews as inputs for the Random Forest classifier that predicts the genre and rating classes.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Random Forests", |
| "sec_num": "5.1.2" |
| }, |
| { |
| "text": "We use dropouts of 0.5, 0.7, and 0.8 at three places in the network to ensure that the model does not overfit. Character embedding based model: Inputs are padded to a length of 300 characters and trimmed if they exceed this length. Character embeddings are generated using an embedding layer that generates 300-dimensional embeddings for each character. We train the model for 300 epochs with a batch size of 512. Word embedding based model: Inputs are padded to a length of 150 words and trimmed if they exceed this length. Word embeddings are generated using an embedding layer that generates 300-dimensional embeddings for each word. We train the model for 200 epochs with a batch size of 512.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "CNN", |
| "sec_num": "5.1.3" |
| }, |
| { |
| "text": "Character-based inputs are padded to a length of 300 and word-based inputs are padded to a length of 100. The embedding layer generates 300dimensional embeddings. The network consists of two LSTM layers followed by multiple Dense layers. A recurrent dropout of 0.4 is used in the first LSTM layer. We train the models for 300 epochs with a batch size of 512.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "LSTM", |
| "sec_num": "5.1.4" |
| }, |
| { |
| "text": "The parameters used are identical to the LSTM parameters except that both the LSTM layers are replaced with GRU layers.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "GRU", |
| "sec_num": "5.1.5" |
| }, |
| { |
| "text": "The model receives a single 300-dimensional sentence embedding that was generated using Doc2Vec. This is passed through a few Dense layers to get our final output. We use dropouts of 0.4 at various places in the network. We train the models for 200 epochs with a batch size of 512.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "FCNN", |
| "sec_num": "5.1.6" |
| }, |
| { |
| "text": "The model as shown in Figure 3 receives three inputs i.e word embedding, char embedding and sentence embedding. The word and char embeddings go into two different LSTM networks. The sentence embeddings go into a fully connected dense network. Each model produces a 100-dimensional output which are concatenated. This 300-dimensional concatenated embedding is given as an input to a dense network to predict the genre and ratings. We use a Dropout of 0.4 throughout the model and train the model for 300 epochs with a batch size of 512.", |
| "cite_spans": [], |
| "ref_spans": [ |
| { |
| "start": 22, |
| "end": 30, |
| "text": "Figure 3", |
| "ref_id": null |
| } |
| ], |
| "eq_spans": [], |
| "section": "Hybrid Model", |
| "sec_num": "5.1.7" |
| }, |
| { |
| "text": "Promising results were obtained in both genre and rating prediction using just the synopsis as the input, as presented in Tables 4 and 5 . For instance, we obtain 91.2% and 90.2% while predicting the Genre and Rating in Telugu respectively.", |
| "cite_spans": [], |
| "ref_spans": [ |
| { |
| "start": 122, |
| "end": 136, |
| "text": "Tables 4 and 5", |
| "ref_id": "TABREF9" |
| } |
| ], |
| "eq_spans": [], |
| "section": "Results", |
| "sec_num": "5.2" |
| }, |
| { |
| "text": "(Dryer, 1997) classified languages in 6 ways depending on whether a subject follows a verb or whether an object follows a verb. The languages we worked on come under SV/OV(Subject-Object-Verb) type of languages. Our dataset consists of multiple languages, some of which are agglutinative (Telugu, Malayalam, Tamil, Japanese and Korean). Our methods obtain good results with various types of languages.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Analysis", |
| "sec_num": "5.3" |
| }, |
| { |
| "text": "Character vs Word Embeddings: We observe that on the whole, word embeddings perform better in general, however in certain cases considering agglutinative languages such as genre prediction in Japanese and rating prediction in Malayalam perform better with character embeddings.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Analysis", |
| "sec_num": "5.3" |
| }, |
| { |
| "text": "Sentence FCNNs: Datasets having small amounts of data work well with sentence vectors. Larger datasets, however, pose issues since the embeddings generated are not precise enough in these cases to differentiate the inputs well.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Analysis", |
| "sec_num": "5.3" |
| }, |
| { |
| "text": "Hybrid Model: LSTM networks learn sequencebased information very well from the character and word embeddings whereas the FCNN learns well from the sentence vectors. Our intuition was that if we develop a model which would use these three models collectively to predict the genre and rating there would be a significant increase in accuracy.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Analysis", |
| "sec_num": "5.3" |
| }, |
| { |
| "text": "So we developed a stacked model which uses combined information from two LSTM networks and one FCNN network to predict the genre and rating. Stacking is an ensemble learning technique and is also known as meta ensembling. This new model outperforms the earlier models as it gives more weight to the individual model where it performs well and gives lesser weight to the individual model where it performs badly. The reason we cannot see a huge change in the rating prediction unlike genre prediction is that the final objective function is not able to learn well from three different models due to the difference in the flow of gradients. If the training is done separately, we would see an increase in accuracy.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Analysis", |
| "sec_num": "5.3" |
| }, |
| { |
| "text": "Traditional vs Deep Learning approaches: There is a huge difference in the performance of traditional machine learning approaches such as SVMs and Random Forests when compared to deep learning based methods such as Sent-FCNN. They all use the same inputs which are the embeddings generated using the Doc2Vec model. We believe that one of the main reasons for this is that deep learning based approaches tend to generalize a lot better as compared to traditional methods and hence they perform a lot better on unseen test data. We have also tried to see how these methods perform if the testing data is just a subset of training data. In this case, we notice that they are both able to represent their training data well and hence achieve similar accuracies during testing. However, this only happens since the number of data points in our dataset are not a lot. When the number of data points increases, deep learning based approaches show tremendous amounts of generalizability which allows them to attain much higher accuracies compared to traditional methods. We validate this by adding noise to our inputs by randomly scaling the inputs up to 15%. This resulted in the failure of traditional approaches as we discussed earlier.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Analysis", |
| "sec_num": "5.3" |
| }, |
| { |
| "text": "Inter-language comparisons: Having a dataset that consists of multiple languages allows us to verify how well our approaches scale and how they can be generalized and applied to data from various domains. MLMRD would also be useful to other researchers who would want to test out their approaches on different languages since multilingual data is becoming popular in recent times. We ob- serve that Telugu, French, Japanese and Korean perform much better than Hindi, Tamil, and Malayalam in rating prediction. This dataset would also allow us to work towards generalized methods that work on multiple forms of inputs that don't require different models to handle different languages which is how traditional approaches work. Qualitative examples with translations and analysis are shown in Appendix A.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Analysis", |
| "sec_num": "5.3" |
| }, |
| { |
| "text": "We provide the multi-lingual dataset MLMRD, consisting of movie genres, ratings and the synopsis which can be used to test various machine learning and NLP based techniques on different kinds of data. We believe that this would be a valuable asset since a lot of these languages are low-resource languages with almost no data available to experiment on. We also propose multiple methods to establish baselines for movie genre and rating prediction based on the synopsis. Additionally, we show how our proposed methods are generalizable and work well on different kinds of data. We plan to extend the approach using movie plots as inputs which would provide us with important information. We also plan to normalize the data collected so that each of the classes have a similar number of data-points.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Conclusion", |
| "sec_num": "6" |
| }, |
| { |
| "text": "32nd Pacific Asia Conference on Language, Information and Computation Hong Kong, 1-3 December 2018Copyright 2018 by the authors", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "", |
| "sec_num": null |
| }, |
| { |
| "text": "32nd Pacific Asia Conference on Language, Information and Computation Hong Kong, 1-3 December 2018Copyright 2018 by the authors", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "", |
| "sec_num": null |
| }, |
| { |
| "text": "32nd Pacific Asia Conference on Language, Information and Computation Hong Kong, 1-3 December 2018Copyright 2018 by the authors", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "", |
| "sec_num": null |
| }, |
| { |
| "text": "32nd Pacific Asia Conference on Language, Information and Computation Hong Kong, 1-3 December 2018Copyright 2018 by the authors", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "", |
| "sec_num": null |
| }, |
| { |
| "text": "32nd Pacific Asia Conference on Language, Information and Computation Hong Kong, 1-3 December 2018Copyright 2018 by the authors", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "", |
| "sec_num": null |
| } |
| ], |
| "back_matter": [ |
| { |
| "text": "A.1 Qualitative example -correct prediction", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "A Appendix", |
| "sec_num": null |
| }, |
| { |
| "text": "Original Synopsis -Translated Synopsis -Plan to make the film on the true story of Flight Attendant Neerja Bhanot was made almost ten years ago. Even after the story was complete, the shooting of the movie was being inhibited due to some reason. Neeraja's Real Life Story is featured in reel Life. Neerja, a resident of Chandigarh, saved 359 passengers on Pan Am Flight 73 in Hijack on September 5, 1986, on the basis of her sense of bravery and she herself became a martyr. This is the first time that the Indian government has given Ashok Chakra only at the age of 23, but Neerja Bhanot was honored. Today, the film was not released on this Friday in Pakistan after the story of the film or some scenes were labeled anti-Pakistan, but the Pakistan government had awarded Neerja the Tamgha-e-Insignias Award. This movie revolves around Neeraja Bhanot (Sonam Kapoor). Along with Neeraja, in this story, there are also mom Rama Bhanot (Shabana Azmi) and Papa (Yogendra Tiku) who are with their daughter at every step. Modeling after study and Neeraja's personal life has also been made part of this story. Actual Genre -Drama Predicted Genre -Drama Actual Rating -3 Predicted Rating -3A.2 Qualitative example -wrong prediction", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "A.1.1 Hindi", |
| "sec_num": null |
| }, |
| { |
| "text": "Original Synopsis -Translated Synopsis -In short, every person thinks low about their lives and live in an other imaginary world. 'Nalo Okadu' is one such story about the life of two people and how it happened to change. Vicky alias Vignesh (Siddharth) works as a torchlight boy in his small theater. He is a good hard working poor guy who feels bad about his looks and status of life. A drug named Lucia brings tremendous changes in his smooth going life. After taking the drug he can live a life that he wants in his dreams. Vicky is a star hero in his dream. He lives a life of star who has a lot of money and fame in the community. The movie is about the parallel lives of Vicky, the one in the theater and the star life and how it changes his life. Actual Genre -Romance Predicted Genre -Comedy Actual Rating -2 Predicted Rating -3", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "A.2.1 Telugu", |
| "sec_num": null |
| } |
| ], |
| "bib_entries": { |
| "BIBREF0": { |
| "ref_id": "b0", |
| "title": "Characterization of movie genre based on music score", |
| "authors": [ |
| { |
| "first": "Austin", |
| "middle": [], |
| "last": "", |
| "suffix": "" |
| } |
| ], |
| "year": 2010, |
| "venue": "Acoustics Speech and Signal Processing (ICASSP), 2010 IEEE International Conference on", |
| "volume": "", |
| "issue": "", |
| "pages": "421--424", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "References [Austin et al.2010] Aida Austin, Elliot Moore, Udit Gupta, and Parag Chordia. 2010. Characterization of movie genre based on music score. In Acoustics Speech and Signal Processing (ICASSP), 2010 IEEE International Conference on, pages 421-424. IEEE. [Basu et al.1998] Chumki Basu, Haym Hirsh, William PACLIC 32", |
| "links": null |
| }, |
| "BIBREF1": { |
| "ref_id": "b1", |
| "title": "Pacific Asia Conference on Language, Information and Computation Hong Kong", |
| "authors": [], |
| "year": 2018, |
| "venue": "", |
| "volume": "", |
| "issue": "", |
| "pages": "", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "Pacific Asia Conference on Language, Information and Computation Hong Kong, 1-3 December 2018 Copyright 2018 by the authors", |
| "links": null |
| }, |
| "BIBREF2": { |
| "ref_id": "b2", |
| "title": "Recommendation as classification: Using social and content-based information in recommendation", |
| "authors": [ |
| { |
| "first": "", |
| "middle": [], |
| "last": "Cohen", |
| "suffix": "" |
| } |
| ], |
| "year": 1995, |
| "venue": "Aaai/iaai", |
| "volume": "20", |
| "issue": "", |
| "pages": "273--297", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "Cohen, et al. 1998. Recommendation as classifica- tion: Using social and content-based information in recommendation. In Aaai/iaai, pages 714-720. [Chung et al.2014] Junyoung Chung, \u00c7 aglar G\u00fcl\u00e7ehre, KyungHyun Cho, and Yoshua Bengio. 2014. Empir- ical evaluation of gated recurrent neural networks on sequence modeling. CoRR. [Cortes and Vapnik1995] Corinna Cortes and Vladimir Vapnik. 1995. Support-vector networks. Machine learning, 20(3):273-297.", |
| "links": null |
| }, |
| "BIBREF3": { |
| "ref_id": "b3", |
| "title": "Movie genre classification using svm with audio and video features", |
| "authors": [ |
| { |
| "first": "Shih-Hao", |
| "middle": [], |
| "last": "Matthew S Dryer ; Yin-Fu Huang", |
| "suffix": "" |
| }, |
| { |
| "first": "", |
| "middle": [], |
| "last": "Wang", |
| "suffix": "" |
| } |
| ], |
| "year": 1997, |
| "venue": "Proceedings of the 8th International Conference on Active Media Technology", |
| "volume": "21", |
| "issue": "", |
| "pages": "1--10", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "Matthew S Dryer. 1997. On the six-way word order typology. Studies in Language. Interna- tional Journal sponsored by the Foundation \"Founda- tions of Language\", 21(1):69-103. [Ghaemmaghami et al.2015] Pouya Ghaemmaghami, Mojtaba Khomami Abadi, Seyed Mostafa Kia, Paolo Avesani, and Nicu Sebe. 2015. Movie genre classification by exploiting meg brain signals. In International Conference on Image Analysis and Processing, pages 683-693. Springer. [Hochreiter and Schmidhuber1997] Sepp Hochreiter and J\u00fcrgen Schmidhuber. 1997. Long short-term memory. Neural Comput., pages 1735-1780. [Huang and Wang2012] Yin-Fu Huang and Shih-Hao Wang. 2012. Movie genre classification using svm with audio and video features. In Proceedings of the 8th International Conference on Active Media Tech- nology, pages 1-10.", |
| "links": null |
| }, |
| "BIBREF4": { |
| "ref_id": "b4", |
| "title": "Distributed representations of sentences and documents", |
| "authors": [ |
| { |
| "first": "Mikolov2014] Quoc", |
| "middle": [], |
| "last": "Le", |
| "suffix": "" |
| }, |
| { |
| "first": "Tomas", |
| "middle": [], |
| "last": "Mikolov", |
| "suffix": "" |
| } |
| ], |
| "year": 2014, |
| "venue": "Proceedings of the 31st International Conference on Machine Learning (ICML-14)", |
| "volume": "", |
| "issue": "", |
| "pages": "", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "and Mikolov2014] Quoc Le and Tomas Mikolov. 2014. Distributed representations of sentences and documents. In Proceedings of the 31st International Conference on Machine Learning (ICML-14), pages 1188-1196. [LeCun et al.2015] Yann LeCun, Yoshua Bengio, and Ge- offrey Hinton. 2015. Deep learning. Nature, pages 436 EP -.", |
| "links": null |
| }, |
| "BIBREF5": { |
| "ref_id": "b5", |
| "title": "Text genre classification with genrerevealing and subject-revealing features", |
| "authors": [ |
| { |
| "first": "Myaeng2002] Yong-Bae", |
| "middle": [], |
| "last": "Lee", |
| "suffix": "" |
| }, |
| { |
| "first": "Sung", |
| "middle": [ |
| "Hyon" |
| ], |
| "last": "Myaeng", |
| "suffix": "" |
| } |
| ], |
| "year": 2002, |
| "venue": "Proceedings of the 25th annual international ACM SIGIR conference on Research and development in information retrieval", |
| "volume": "", |
| "issue": "", |
| "pages": "145--150", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "and Myaeng2002] Yong-Bae Lee and Sung Hyon Myaeng. 2002. Text genre classification with genre- revealing and subject-revealing features. In Proceed- ings of the 25th annual international ACM SIGIR con- ference on Research and development in information retrieval, pages 145-150. ACM.", |
| "links": null |
| }, |
| "BIBREF6": { |
| "ref_id": "b6", |
| "title": "Random forest: a classification and regression tool for compound classification and qsar modeling", |
| "authors": [ |
| { |
| "first": "Bastiaan", |
| "middle": [], |
| "last": "Quast. ; Zeeshan Rasheed", |
| "suffix": "" |
| }, |
| { |
| "first": "Mubarak", |
| "middle": [], |
| "last": "Shah", |
| "suffix": "" |
| }, |
| { |
| "first": ";", |
| "middle": [], |
| "last": "Gabriel", |
| "suffix": "" |
| }, |
| { |
| "first": "S", |
| "middle": [], |
| "last": "Sim\u00f5es", |
| "suffix": "" |
| }, |
| { |
| "first": "Jonatas", |
| "middle": [], |
| "last": "Wehrmann", |
| "suffix": "" |
| }, |
| { |
| "first": "C", |
| "middle": [], |
| "last": "Rodrigo", |
| "suffix": "" |
| }, |
| { |
| "first": "Duncan D Ruiz ; Vladimir", |
| "middle": [], |
| "last": "Barros", |
| "suffix": "" |
| }, |
| { |
| "first": "Andy", |
| "middle": [], |
| "last": "Svetnik", |
| "suffix": "" |
| }, |
| { |
| "first": "Christopher", |
| "middle": [], |
| "last": "Liaw", |
| "suffix": "" |
| }, |
| { |
| "first": "Christopher", |
| "middle": [], |
| "last": "Tong", |
| "suffix": "" |
| }, |
| { |
| "first": "Robert", |
| "middle": [ |
| "P" |
| ], |
| "last": "Culberson", |
| "suffix": "" |
| }, |
| { |
| "first": "", |
| "middle": [], |
| "last": "Sheridan", |
| "suffix": "" |
| }, |
| { |
| "first": "", |
| "middle": [], |
| "last": "Bradley P Feuston", |
| "suffix": "" |
| } |
| ], |
| "year": 2002, |
| "venue": "Neural Networks (IJCNN), 2016 International Joint Conference on", |
| "volume": "2", |
| "issue": "", |
| "pages": "973--982", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "Bastiaan Quast. 2016. rnn: a recurrent neu- ral network in r. Working Papers. [Rasheed and Shah2002] Zeeshan Rasheed and Mubarak Shah. 2002. Movie genre classification by exploiting audio-visual features of previews. In Pattern Recog- nition, 2002. Proceedings. 16th International Confer- ence on, volume 2, pages 1086-1089. [Rehurek and Sojka2010] Radim Rehurek and Petr So- jka. 2010. Software framework for topic modelling with large corpora. In THE LREC 2010 WORKSHOP ON NEW CHALLENGES FOR NLP FRAMEWORKS, pages 45-50. [Sim\u00f5es et al.2016] Gabriel S Sim\u00f5es, Jonatas Wehrmann, Rodrigo C Barros, and Duncan D Ruiz. 2016. Movie genre classification with con- volutional neural networks. In Neural Networks (IJCNN), 2016 International Joint Conference on, pages 259-266. IEEE. [Svetnik et al.2003] Vladimir Svetnik, Andy Liaw, Christopher Tong, J Christopher Culberson, Robert P Sheridan, and Bradley P Feuston. 2003. Random for- est: a classification and regression tool for compound classification and qsar modeling. Journal of chemical information and computer sciences, 43(6):1947-1958. [Wehrmann and Barros2017] J\u00f4natas Wehrmann and Ro- drigo C Barros. 2017. Movie genre classification: A multi-label approach based on convolutions through time. Applied Soft Computing, 61:973-982.", |
| "links": null |
| }, |
| "BIBREF7": { |
| "ref_id": "b7", |
| "title": "Guizhong Liu, and Andrew Perkis. 2010. A semantic framework for video genre classification and event analysis", |
| "authors": [ |
| { |
| "first": "Chin-Chia Michael", |
| "middle": [], |
| "last": "Yeh", |
| "suffix": "" |
| }, |
| { |
| "first": "Yi-Hsuan", |
| "middle": [], |
| "last": "Yang", |
| "suffix": "" |
| } |
| ], |
| "year": 2010, |
| "venue": "Proceedings of the 2nd ACM International Conference on Multimedia Retrieval", |
| "volume": "25", |
| "issue": "", |
| "pages": "747--750", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "and Yang2012] Chin-Chia Michael Yeh and Yi- Hsuan Yang. 2012. Supervised dictionary learning for music genre classification. In Proceedings of the 2nd ACM International Conference on Multimedia Re- trieval, page 55. ACM. [You et al.2010] Junyong You, Guizhong Liu, and An- drew Perkis. 2010. A semantic framework for video genre classification and event analysis. Signal Pro- cessing: Image Communication, 25(4):287-302. [Zhou et al.2010] Howard Zhou, Tucker Hermans, As- mita V Karandikar, and James M Rehg. 2010. Movie genre classification via scene categorization. In Pro- ceedings of the 18th ACM international conference on Multimedia, pages 747-750.", |
| "links": null |
| }, |
| "BIBREF8": { |
| "ref_id": "b8", |
| "title": "Pacific Asia Conference on Language, Information and Computation Hong Kong", |
| "authors": [], |
| "year": 2018, |
| "venue": "", |
| "volume": "", |
| "issue": "", |
| "pages": "", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "Pacific Asia Conference on Language, Information and Computation Hong Kong, 1-3 December 2018 Copyright 2018 by the authors", |
| "links": null |
| } |
| }, |
| "ref_entries": { |
| "FIGREF0": { |
| "type_str": "figure", |
| "text": "Branched CNN based genre/rating prediction model architecture with character inputs which considers the synopsis to perform the prediction. The Conv blocks represented in the figure consist of Convolution, ReLU and Max-pool layers. We used Dropouts at various places for regularization. Word inputs are similar except that the input size is different.", |
| "uris": null, |
| "num": null |
| }, |
| "FIGREF1": { |
| "type_str": "figure", |
| "text": "LSTM based genre/rating prediction model architecture with character inputs which considers the synopsis to perform the predictions. Replacing LSTM cells with GRUs would give us GRU models Recurrent networks: For LSTM(Hochreiter and Schmidhuber, 1997),GRU (Chung et al., 2014) andRNN (Quast, 2016) based networks as shown in", |
| "uris": null, |
| "num": null |
| }, |
| "TABREF1": { |
| "content": "<table/>", |
| "num": null, |
| "html": null, |
| "text": "", |
| "type_str": "table" |
| }, |
| "TABREF5": { |
| "content": "<table/>", |
| "num": null, |
| "html": null, |
| "text": "Preprocessing the genres to form 9 genre classes that are used for genre prediction.", |
| "type_str": "table" |
| }, |
| "TABREF7": { |
| "content": "<table/>", |
| "num": null, |
| "html": null, |
| "text": "", |
| "type_str": "table" |
| }, |
| "TABREF9": { |
| "content": "<table><tr><td>Model</td><td colspan=\"8\">Train % Telugu Hindi Tamil Malayalam French Japanese Korean</td></tr><tr><td>Char-CNN</td><td/><td>85.2</td><td>83.1</td><td>82.7</td><td>82.1</td><td>89.1</td><td>90.5</td><td>90.4</td></tr><tr><td>Word-CNN</td><td/><td>82.2</td><td>80.1</td><td>83.2</td><td>81.1</td><td>89.0</td><td>89.4</td><td>89.8</td></tr><tr><td>Concat-CNN</td><td/><td>90.2</td><td>89.5</td><td>89.2</td><td>89.5</td><td>89.2</td><td>91.5</td><td>89.8</td></tr><tr><td>Char-LSTM</td><td>80</td><td>85.9</td><td>81.7</td><td>82.2</td><td>82.1</td><td>89.1</td><td>89.1</td><td>93.5</td></tr><tr><td>Word-LSTM</td><td/><td>86.2</td><td>83.5</td><td>80.4</td><td>83.2</td><td>88.9</td><td>89.0</td><td>91.4</td></tr><tr><td>Concat-LSTM</td><td/><td>89.0</td><td>88.9</td><td>88.9</td><td>89.2</td><td>88.9</td><td>91.2</td><td>93.5</td></tr><tr><td>Char-GRU</td><td/><td>85.8</td><td>82.4</td><td>81.2</td><td>81.1</td><td>89.2</td><td>89.5</td><td>91.2</td></tr><tr><td>Word-GRU</td><td/><td>86.1</td><td>83.2</td><td>81.1</td><td>80.5</td><td>89.0</td><td>89.1</td><td>91.7</td></tr><tr><td>Concat-GRU</td><td/><td>88.9</td><td>89.5</td><td>88.9</td><td>89.5</td><td>88.9</td><td>91.2</td><td>90.7</td></tr><tr><td>Sent-FCNN</td><td/><td>85.2</td><td>80.9</td><td>84.2</td><td>80.2</td><td>89.2</td><td>89.5</td><td>91.3</td></tr><tr><td>Concat-FCNN</td><td/><td>80.0</td><td>78.0</td><td>89.1</td><td>69.5</td><td>73.1</td><td>77.9</td><td>73.3</td></tr><tr><td>Hybrid-model</td><td/><td>84.6</td><td>83.1</td><td>81.0</td><td>80.0</td><td>80.5</td><td>80.8</td><td>82.5</td></tr><tr><td>SVM Random Forests</td><td>80</td><td>58.0 68.8</td><td>50.9 52.9</td><td>45.9 54.0</td><td>47.3 44.8</td><td>44.7 43.1</td><td>48.8 55.5</td><td>45.8 54.0</td></tr></table>", |
| "num": null, |
| "html": null, |
| "text": "Genre prediction accuracies for various models on each of the languages in MLMRD.", |
| "type_str": "table" |
| }, |
| "TABREF10": { |
| "content": "<table/>", |
| "num": null, |
| "html": null, |
| "text": "Rating prediction accuracies for various models on each of the languages in MLMRD.", |
| "type_str": "table" |
| } |
| } |
| } |
| } |
