Fine‑grained Character Sentiment Analysis in Scripts: Models, Challenges, and Future Directions

Script analysis is the first link in the content production chain. By combining expert knowledge, big data, and natural language processing (NLP) techniques, iQIYI helps business units quickly evaluate scripts. Character sentiment analysis is a crucial task in this pipeline.

2. Common Sentiment Analysis Tasks

2.1 Traditional Sentiment Analysis usually predicts a single polarity (positive, negative, neutral) for a sentence or document. Methods include lexical‑rule analysis, algorithmic models (NB, LR, SVM, CNN, RNN), and hybrid approaches that combine dictionaries with supervised models.

2.2 Target‑dependent Sentiment Analysis predicts sentiment toward a specific target within a sentence. Example: "Zhang San is popular, but Li Si is not." The sentiment toward Zhang San is positive, while that toward Li Si is negative. Deep learning models such as TD‑LSTM and TC‑LSTM are introduced.

TD‑LSTM uses two LSTMs to model the context before and after the target word:

TC‑LSTM extends TD‑LSTM by concatenating the target embedding with each word embedding, allowing the model to capture target‑context interactions.

2.3 Aspect‑level Sentiment Analysis deals with abstract aspects that may not appear explicitly in the text. Models such as AT‑LSTM, ATAE‑LSTM, and TNET are described.

AT‑LSTM adds an attention mechanism to a Bi‑LSTM, using aspect embeddings to compute attention weights for each hidden state.

ATAE‑LSTM concatenates aspect embeddings with word embeddings at both the input and hidden layers, further enhancing aspect‑context interaction.

TNET employs a Bi‑LSTM encoder, multiple CPT (Context‑Preserving Transformation) layers, a convolutional layer, and a final Softmax classifier.

3. Character‑level Fine‑grained Sentiment Analysis

Challenges include (1) recognizing character names (NER), (2) building a business‑specific sentiment dimension model, and (3) deep semantic understanding of long scripts.

For NER, traditional HMM/CRF models are insufficient; deep models such as BiLSTM+CRF and LatticeLSTM are employed, combined with unsupervised “new‑word discovery” to generate training data.

The first‑version sentiment model uses a ten‑emotion wheel (love, joy, trust, anticipation, surprise, doubt, worry, anger, sorrow, fear) to label each character’s emotion per line.

Model architecture for action description combines BERT (for feature extraction) with TD‑LSTM (target‑dependent analysis). For dialogue, a BERT + multi‑layer CNN architecture is used to capture shorter, more direct sentiment cues.

The second‑version model introduces a global GRU to capture long‑range context and an emotion‑GRU per character to record cumulative emotional states. Self‑attention encodes the whole preceding context, and updates are performed for each character present in the current sentence.

3.4 Case Study – The Korean film “Parasite” script is used to compare the two models. The first version predicts a mild sadness for the character Kim Ki‑taek, while the second version, leveraging longer context and emotion history, predicts high happiness, confidence, and anticipation, which aligns better with the narrative.

3.5 Summary and Future Work

The current fine‑grained character sentiment analysis system is deployed in iQIYI’s script evaluation pipeline, aiding content teams in character development, conflict detection, and plot pacing. Future improvements include incorporating genre‑specific knowledge, handling artistic devices (flashbacks, foreshadowing), simplifying the network architecture, and exploring sentence‑matching formulations to better exploit pretrained models.

4. References

[1] Duyu Tang, Bing Qin, Xiaocheng Feng, Ting Liu. 2016. Effective LSTMs for Target‑Dependent Sentiment Classification.

[2] Yequan Wang, Minlie Huang, Li Zhao, Xiaoyan Zhu. 2016. Attention‑based LSTM for Aspect‑level Sentiment Classification.

[3] Xin Li, Lidong Bing, Wai Lam, Bei Shi. 2018. Transformation Networks for Target‑Oriented Sentiment Classification.

[4] Navonil Majumder et al. 2019. DialogueRNN: An Attentive RNN for Emotion Detection in Conversations.

[5] Yue Zhang and Jie Yang. 2018. Chinese NER Using Lattice LSTM.

[6] Liang‑Chih Yu et al. 2016. Building Chinese Affective Resources in Valence‑Arousal Dimensions.