How Dynamic Template Matching Transforms User Review Tag Extraction

Background

Online shopping and travel bookings increasingly rely on user reviews, and extracting meaningful tags from these reviews is crucial for helping users make decisions and for improving platform conversion rates.

Problems with Existing Tagging Methods

Preset tags : Fixed tags are limited in number and often mismatch user content.

Syntactic analysis : Generates many tags for large volumes of reviews, causing high computational cost and maintenance difficulty.

Multi‑level tag definition : Produces massive maintenance work and lacks flexibility, with tags often being keyword‑plus‑indicator combinations that do not reflect natural user language.

Dynamic Tag Extraction via Template Matching

To address these issues, the team proposes a flexible, dynamic method that matches predefined sentence templates to user comments, mapping each template to a fixed tag category while allowing the template to consist of multiple word groups, thereby reducing the number of preset tags and better aligning with user language habits.

Key Concepts

Tag : A specific description of a piece of information, e.g., “near Beijing subway station”.

Sentence pattern : A collection of similar tags, representing an “evaluation way”.

Tag category : A group of sentence patterns that share a common evaluation theme.

A tag category contains m sentence patterns; each pattern can generate n tags, so a category may correspond to up to m×n tags.

System Architecture

The system consists of two main parts: definition of sentence patterns and automated generation of those patterns. The diagram below shows the overall structure.

Step 1: Build Sentence Library

A sentence library is the collection of all predefined sentence templates. The following figure illustrates its layout.

Step 2: Build Word Library

The word library contains word groups and the words belonging to each group. Each word group has a unique identifier and summarizes its words. Examples include a group for “shuttle bus” (words: shuttlebus, 班车, etc.) and a group for “near” (words: near, close, 1 minute walk, etc.).

Step 3: Classify Sentence Patterns into Tag Categories

Sentence patterns are grouped into tag categories; for example, the category “service good” includes patterns like {boss}{enthusiastic} and {reception}{professional}. All tags generated from these patterns belong to the same category, though the concrete wording may differ.

Step 4: Combine Word Groups to Form Sentences

Each sentence pattern is a logical expression composed of word groups. For example, the pattern {provide}[{subway}OR{pier}OR{bus stop}OR{train station}OR{airport}OR{city center}]{shuttle} combines a normal group (“provide”, “shuttle”) with an independent group (e.g., “subway”). Independent groups and POI groups are displayed separately when matched.

Step 5: Sentence Matching and Tag Generation

UGC reviews are split into clauses. Matching proceeds sequentially: each word group is matched in order, respecting the position of previously matched words. If all groups find a matching word, the combined words form a tag. Rules such as sequential matching, distance thresholds, and negation detection are applied to avoid incorrect matches.

Matching Rules

Sequential matching : Ensures the order of word groups is respected (e.g., “airport has shuttle to hotel” vs. “hotel has shuttle to airport”).

Distance threshold : If the distance between matched words exceeds a preset limit, the match is discarded.

Negation handling : Tags are rejected when a negation word appears in the clause.

Confusion word library : Words that are easily confused (e.g., “好像”) are checked to prevent false matches.

Determine Display Tags

For each tag category, the most frequent tags generated from its patterns are selected as the displayed tag. If a pattern contains independent words that must be shown separately, its top‑frequency tag is displayed independently.

Unmatched Sentences Processing

Clauses that fail to match any pattern are sent to an automatic sentence generation pipeline that uses content classification, syntactic and dependency analysis, and semantic analysis to propose new patterns and word groups, which can then be reviewed and added to the libraries.

Conclusion

The template‑matching approach dramatically reduces the number of preset tags while producing tags that align with natural user language, improving recall and relevance of extracted information. Future work will cover automatic sentence generation.