InferSent

InferSent is a supervised learning framework for producing general-purpose, fixed-length representations of sentences. Developed by researchers at Facebook AI Research (FAIR), it aims to learn universal sentence representations that can be used across diverse NLP tasks.

Method: It learns sentence embeddings by training a sentence encoder to predict the semantic relationship between two sentences. The encoder is a bidirectional LSTM that processes each sentence and produces a fixed-length vector via max-pooling over time. The two sentence representations are then combined and passed to a classifier that predicts entailment, contradiction, or neutrality. The model is trained on natural language inference data, typically using SNLI and MultiNLI datasets.

Usage: After training, the fixed-length sentence vector can be used as generic features for downstream tasks. Researchers and practitioners commonly apply simple classifiers (e.g., logistic regression) on top of the embeddings for tasks such as sentiment analysis, question classification, and semantic similarity, often achieving competitive results with less task-specific tuning. The authors released pre-trained InferSent embeddings that can be used directly with standard NLP toolchains.

Variants and data: The original work emphasizes training on NLI data to obtain transferable representations and provides pre-trained models based on SNLI (and later expanded to MultiNLI). The approach is notable for showing that supervised signals from NLI can yield robust sentence representations.

Limitations: Dependence on NLI data and domain mismatch can limit applicability to languages or domains not represented in the training data; as with other neural representations, interpretability remains challenging.