Neural Aspect and Opinion Term Extraction with Mined Rules as Weak Supervision

• Aspect and opinion term extraction targets product attributes and reviewer sentiments; manual rules are labor-intensive and natural-language variability limits their performance.Neural sequence-labeling approaches can perform better but need substantial labeled data to reach their potential.Source: §1 · p. 1
• RINANTE mines rules from dependency parsing and trains a BiLSTM-CRF using rule-labeled auxiliary reviews alongside human annotations.The rules provide weak supervision, while human annotations serve as ground-truth supervision.Source: §1 · pp. 1–2
• Training with human-annotated data and rule-annotated data significantly improves neural-model performance on three SemEval datasets.The introduction reports a significant improvement but gives no numerical results here.Source: §1 · pp. 1–2

• Dependency-based rule methods can extract patterned terms, but manual design is labor-intensive and rules involving up to three words limit flexibility.Liu et al. (2015b) select from previously designed rules but do not discover rules automatically.Source: §2 · p. 2
• Topic models provide coarse-grained aspects and related words, but do not extract exact aspect terms from review sentences.Source: §2 · p. 2
• Learning-based methods label words with BIO tags and use CRFs; neural variants include dependency-tree models and CNNs.Word embeddings and handcrafted features are also used.Source: §2 · p. 2
• Semi-supervised sequence labeling and transfer learning use extra data, but RINANTE instead uses rule outputs for the same problem rather than labels from a related task.The approach is also related to weak labeling and distant supervision.Source: §2 · p. 2

3 RINANTE

• RINANTE mines aspect and opinion rules from labeled data, applies them to auxiliary reviews as weak labels, then trains a neural model with both weakly and manually labeled data.The labeled and auxiliary reviews are for the same or similar product types; the auxiliary set is usually larger.Source: §3 · pp. 2–3

3.1 Rule Mining Algorithm

• Candidate rules use dependency relations involving labeled aspect terms, generalize word patterns by part of speech or opinion vocabulary, and include patterns based on one or two relations.The algorithm limits mined rules to patterns involving no more than three words because larger patterns may offer little recall and cost more to mine and validate.Source: §3.1 · pp. 3–4
• The algorithm retains dependency-pattern candidates occurring more than a predefined frequency threshold T, then merges candidates based on single relations and relation pairs.T can be set based on the total number of sentences in the training set.Source: §3.1 · p. 4
• Candidate rules are evaluated on a validation set and filtered by a precision threshold; matched seeds are expanded into verbs or consecutive-noun phrases, with a filter for known incorrect phrases.A dictionary of frequent training-set aspect terms is also used for direct matching.Source: §3.1 · p. 4
• Opinion-rule mining follows a similar procedure, but does not generate patterns tied to an opinion-word vocabulary and considers adjectives, nouns, and verbs as possible opinion terms.Source: §3.1 · p. 4

3.2 Neural Model

• Because rule labels can assign a word to both term types, the model uses separate aspect-rule and opinion-rule tag sequences alongside a manual-label prediction task.The three prediction tasks are denoted ta, to, and tm, respectively.Source: §3.2 · pp. 4–5
• Both neural models use pretrained word embeddings and BiLSTM-CRF prediction; the Shared BiLSTM Model shares one BiLSTM across tasks, while the Double BiLSTM Model uses two task-specific BiLSTMs.In the Double BiLSTM Model, the manual-label task receives concatenated outputs from both BiLSTMs.Source: §3.2 · p. 5
• The models are trained either by alternating among the three tasks or by pretraining on rule-derived tasks before training on manual labels, with early stopping based on validation F1 scores.For pretraining, stopping uses the sum of the aspect-rule and opinion-rule F1 scores; later training uses the sum of aspect and opinion F1 scores for manual labels.Source: §3.2 · p. 5

4.1 Datasets

• The three evaluation datasets are SemEval-2014 Restaurants, SemEval-2014 Laptops, and SemEval-2015 Restaurants; Yelp and Amazon reviews supply auxiliary data.The Yelp data supports the restaurant datasets, while Amazon Electronics supports the laptop dataset.Source: §4.1 · pp. 5–6, Table 1 · p. 6

4.2 Experimental Setting

• For SE14-L, rules label 156,014 Amazon review sentences; for SE14-R and SE15-R, they label 913,443 sampled Yelp sentences.The authors use Stanford CoreNLP for dependency parsing and POS tagging, with frequency threshold 10 and precision threshold 0.6.Source: §4.2 · p. 6

4.3 Performance Comparison

• All four RINANTE variants outperform RINANTE (no rule), with especially significant gains on SE14-L and SE15-R.Table 2 reports F1 scores for aspect and opinion term extraction across approaches.Source: Table 2 · p. 7, §4.3 · p. 7
• RINANTE-Double-Pre is more stable overall and leads on SE14-L and SE15-R, while RINANTE-Shared-Alt is slightly better on SE14-R.The authors suggest the separate BiLSTM layers work more stably; both training methods can yield good performance.Source: §4.3 · pp. 7–8
• The mined rules outperform Double Propagation because the algorithm mines hundreds of effective rules, compared with its eight manually designed rules.Source: §4.3 · p. 8
• Excluding RINANTE-Double-Pre†, RINANTE misses the best aspect-term result only on SE14-L and SE15-R; DE-CNN leads on SE14-L, while † is best on SE15-R.The comparison notes that DE-CNN and HAST focus only on aspect terms, whereas this approach extracts both aspect and opinion terms.Source: Table 2 · p. 7, §4.3 · p. 8

4.4 Mined Rule Results

• Rule mining takes under 10 seconds per dataset; SE15-R has the fewest rules because it has the fewest training samples, resulting in inferior rule performance.Source: §4.4 · p. 8

4.5 Case Study

• Case examples show rules can recover an obvious aspect missed without them, while RINANTE can extend a rule’s “microphones” to the correct phrase “external microphones.”The examples also show mined rules can help avoid incorrect term extractions.Source: §4.5 · p. 8

5 Conclusion and Future Work

• The approach mines dependency-relation rules, applies them to unlabeled data, and combines those annotations with a small human-labeled set to train neural models.Source: §5 · p. 8
• Table 5 presents example sentences and aspect terms extracted by different approaches, marking correct terms in bold and using “-” for no extraction.Source: Table 5 · p. 9
• The authors plan to apply the approach’s main idea to other tasks.Source: §5 · p. 9

A Experimental Results with BERT

• BERT fine-tuning feeds each token’s final hidden representation to a classification layer, while feature-based BERT supplies pretrained Transformer hidden layers to a sequence-labeling model.The feature-based approach can also be used by RINANTE.Source: A Experimental Results with BERT · p. 10
• The comparison includes word2vec BiLSTM-CRF, BERT fine-tuning, BERT feature-based, and RINANTE+BERT using the top four hidden layers as features.Both BERT feature-based and RINANTE+BERT use those layers as input features.Source: A Experimental Results with BERT · p. 10
• The experiments use pretrained BERT-Base, Uncased, further pretrained on Yelp reviews for restaurant data and Amazon reviews for laptop data, with 200-dimensional BiLSTMs.Source: A Experimental Results with BERT · p. 10
• BERT improves performance over word2vec, and RINANTE further improves performance with BERT contextual embeddings.Table 6 reports aspect and opinion term extraction performance.Source: A Experimental Results with BERT · p. 10, Table 6 · p. 11