Traditional sentiment analysis treats a sentence as a single emotional unit, assigning it a scalar value from "negative" to "positive." While this works for simple statements, it fails on the nuanced feedback common in real-world reviews. Consider the sentence: "The screen is stunning, but the battery life is disappointing." A global classifier might label this as "neutral" by averaging the two polarities, effectively hiding two critical, opposing pieces of information.
To solve this, researchers have moved toward Aspect-Based Sentiment Analysis (ABSA). The goal is to disentangle the sentiment directed at specific targets (like "screen" or "battery") within the same context, ensuring that the praise for one doesn't "leak" into the criticism of another.
The Problem of Sentiment Leakage
The primary challenge in multi-target sentences is sentiment leakage. When a model processes a sentence, the emotional weight of a strong adjective (like "stunning") can overwhelm the representation of the entire sentence. If the model isn't careful, it might incorrectly attribute that positivity to every entity mentioned.
Hao Wang, Bin Liang, and their colleagues identified this as a struggle between global and local contexts. In their work on Gated Hierarchical LSTMs, they argue that models must simultaneously maintain a "global" view of the sentence while isolating a "local" region around the target. By using a gated merge operation, the model can decide how much of the global sentiment is actually relevant to the specific target at hand.
Dual Attention and Contextual Noise
Attention mechanisms are the standard tool for focusing on relevant words, but even they can be distracted by "contextual noise"—irrelevant sentiment words belonging to other targets. Jiachen Ding and Weidong Qiu addressed this in ETAN by employing a dual-attention mechanism. One layer of attention looks at the whole sentence, while a second layer focuses strictly on the words immediately surrounding the target. This prevents the model from being "blinded" by a positive word at the end of a sentence when it is trying to evaluate a negative target at the beginning.
Similarly, Feifan Fan and Yansong Feng proposed the Multi-grained Attention Network (MGAN), which introduces an "aspect alignment loss." This explicitly penalizes the model if it uses the same attention weights for two different targets with different sentiments. By forcing the model to "partition" the sentence, it learns to treat different entities as independent emotional subjects.
Disentangling Features with Capsule Networks
When sentiments overlap in complex ways, simple attention might not be enough to separate the features. Chunning Du and Jianxin Liao turned to Capsule Networks to solve this. In IACapsNet, they use EM routing to cluster features that belong together. Unlike standard pooling, which might blur features, capsules can maintain the "pose" or specific properties of a sentiment, making it easier to route the "disappointing" feature to "battery" and the "stunning" feature to "screen."
This need for robust separation led Qingnan Jiang and Lei Chen to develop the MAMS (Multi-Aspect Multi-Sentiment) dataset. They observed that many existing benchmarks were too easy because they often contained only one sentiment per sentence. MAMS forces models to handle sentences where every instance has at least two different aspects with conflicting polarities, providing a true test of a model's ability to resolve fine-grained relationships.
Structural and Latent Scopes
Beyond attention and capsules, some researchers look at the underlying structure of language to isolate sentiment. Thien Hai Nguyen and Kiyoaki Shirai developed PhraseRNN, which uses dependency trees to create "structural isolation." By following the grammatical links between words, the model can see that "slow" is syntactically connected to "service" even if it is physically closer to "food" in the sentence.
Taking this further, Hao Li and Wei Lu introduced Latent Sentiment Scopes. Instead of assuming a fixed window of words around a target, they treat the "scope" of a sentiment as a hidden variable. This allows the model to dynamically learn how far a sentiment's influence travels, capturing long-distance modifiers that traditional fixed-window models often miss.
Why It Matters: From Scores to Insights
The shift from global sentiment to disentangled aspect-level analysis is more than a technical hurdle; it is a shift toward actionable data. As noted in Social Media Users’ Perceptions of a Wearable Mixed Reality Headset, a company doesn't just need to know if people "like" their product. They need to know that users love the immersion but hate the cost. By resolving these conflicting signals within complex user feedback, these models provide the granularity required to understand the specific drivers of human opinion.