When a sentence contains a single opinion about a single topic—such as "The food was spectacular"—sentiment analysis is straightforward. However, natural language is rarely so simple. Reviews, blog posts, and news articles frequently evaluate multiple targets or aspects within the same sentence, often with sharply contrasting emotions: "The food was spectacular, but the service was absolutely abysmal."
For standard deep learning models, these multi-target sentences present a severe challenge. When a global sentence encoder compresses this text into a single vector representation, the positive features of "spectacular" and the negative features of "abysmal" bleed into one another. This phenomenon is known as sentiment leakage or feature interference.
As Du et al. (2019) observe, sentences with multiple targets and conflicting sentiments lead to overlapping feature representations in neural networks, making it incredibly difficult for classifiers to correctly assign specific polarities to the appropriate aspect terms. If a model cannot effectively isolate the relevant context for one target from the opinions directed at others, a dominant sentiment regarding one feature can easily overshadow a more nuanced or opposing sentiment for a secondary feature (Ji et al. (2020)).
To solve this, researchers have developed architectures designed to partition context and disentangle overlapping features, ensuring that each target is evaluated only by the words that logically belong to it.
Context Partitioning via Gating and Co-Attention
The most direct way to prevent sentiment leakage is to build a dynamic wall between different parts of a sentence. Rather than allowing every word to influence the representation of every target, models can use gating mechanisms and multi-grained attention to filter out irrelevant context.
Xue et al. (2018) pioneered this approach using Gated Convolutional Networks (GCAE). Instead of relying on complex recurrent layers, they introduced a specialized gating unit that receives both the sentence context and the specific target aspect. The gate acts as a selective valve: it blocks features that are highly charged with sentiment but belong to a different aspect, allowing only the target-relevant emotional signals to pass through.
Similarly, Wang et al. (2018) developed a Gated Hierarchical LSTM (GH-LSTM) that simultaneously tracks a local "region" immediately surrounding a target and the global context of the entire sentence. By using a gated merge operation, the model balances these two perspectives, ensuring that global sentiment words do not dilute the specific local descriptors of a target.
To prevent models from treating multiple targets as completely isolated entities, Fan et al. (2018) introduced a Multi-grained Attention Network (MGAN) featuring an "aspect alignment loss." This loss explicitly forces the model to contrast and partition contextual information across different targets within the same sentence, teaching the network how to actively differentiate between conflicting polarities rather than just ignoring the noise.
Syntactic Disentanglement
Physical proximity in a sentence does not always equal logical connection. In the phrase "The screen, which took months to ship, is beautiful," the word "beautiful" is physically closer to "ship" than to "screen," yet it describes the screen. When multiple targets are present, sequential models often get confused by this physical distance.
To fix this, researchers leverage the grammatical structure of the sentence using dependency graphs. By transforming a sentence into a tree of syntactic relationships (e.g., mapping adjectives directly to the nouns they modify), the structural distance between a target and its true descriptor becomes much shorter than their physical distance in the text.
[The food] <---(nsubj)--- [was spectacular]
|
(conj)
v
[the service] <---(nsubj)--- [was abysmal]
Nguyen et al. (2015) first demonstrated the power of this concept by combining constituent (phrase-based) and dependency (relation-based) trees within a recursive neural network (PhraseRNN), creating a richer representation that prevents the emotional polarity of one aspect from leaking into another.
Building on this, Bai et al. (2020) introduced the Relational Graph Attention Network (RGAT). Instead of treating all syntactic connections equally, RGAT explicitly incorporates typed dependency labels (such as "subject" or "adjective modifier") into a graph attention framework. This allows the model to precisely map the logical flow of descriptors to their respective entities, effectively insulating the target from nearby, unrelated sentiment words.
More recently, Meng et al. (2022) utilized a Multi-Task Dual-Tree Network (MTDTN) that deploys constituency trees to find target boundaries and modified dependency trees to map relationships, ensuring that conflicting sentiment polarities are cleanly separated.
Robust Aggregation: Capsules and Multi-Instance Learning
Another way to combat feature interference is to change how the model aggregates information. Standard neural networks often use global pooling layers that average word vectors together, which inherently blurs conflicting signals.
To preserve the distinct identity of conflicting sentiments, Du et al. (2019) replaced standard pooling with a capsule network (IACapsNet) that uses Expectation-Maximization (EM) routing. Instead of forcing all features into a single vector, the EM routing algorithm clusters overlapping features into distinct "capsules" based on their similarity and relevance to specific target aspects. This prevents a strong positive sentiment from bleeding into a weak negative capsule.
Alternatively, researchers have reframed aspect sentiment analysis as a Multi-Instance Learning (MIL) problem. Li et al. (2020) developed AC-MIMLLN, a model that treats individual words as independent "instances" within a bag, rather than fusing them into a global sentence representation. By calculating sentiment as an aggregation of these discrete word-level instances, the model avoids the diluted representations that occur when global sentence encoders attempt to force conflicting emotional signals into a single space.
Reformulating the Problem: Spans and Queries
Finally, some researchers argue that the traditional sequence-labeling approach—where models assign a sentiment label to each individual token—is fundamentally flawed for multi-target text. As Huang et al. (2019) point out, token-level tagging often produces highly inconsistent sentiment predictions across a single multi-word entity (e.g., labeling "battery" as positive and "life" as negative).
To maintain semantic unity, Huang et al. (2019) and Bing et al. (2019) proposed span-based joint models. By treating target phrases as single, unbroken spans rather than isolated words, the model enforces sentiment consistency across the entire entity. Wang et al. (2018) integrated a Conditional Random Field (CRF) layer to identify these latent opinion expressions as coherent spans, providing clear, human-readable justifications for why a sentiment belongs to a specific target.
Taking this a step further, Yu et al. (2021) reframed aspect-level sentiment analysis as a Machine Reading Comprehension (MRC) task. Instead of scanning a sentence linearly, the model treats the target aspect as a question (e.g., "What is the sentiment towards the food?") and the sentence as the context. By turning the task into a query-driven search, the model naturally ignores the text segments that answer questions about other targets, completely bypassing the problem of feature interference.