Building and Deploying Language Models for Text Quality Evaluation and Generation

Language models are fundamental probabilistic tools in natural language processing, widely used in tasks such as speech recognition, sentiment analysis, and machine translation, and they can improve the ranking of rental and second‑hand housing posts by assessing textual quality.

The development of language models has progressed from n‑gram approaches to recurrent neural networks (RNN), long short‑term memory (LSTM), gated recurrent units (GRU) and their variants, as illustrated in the evolution diagram.

Training a language model involves three main steps: data preparation (collecting relevant posts, cleaning, removing irrelevant content, handling punctuation, stop‑words, HTML tags, and encoding words with one‑hot or embedding vectors), model training (using cross‑entropy loss, optimizers such as Adam, hyper‑parameters like batch_size, num_steps, hidden_size, dropout, early stopping, and evaluating with perplexity), and finally model export.

For deployment, TensorFlow provides several formats: checkpoint (weights only), GraphDef (graph structure without weights), and SavedModel (combined graph and weights) which is used with TensorFlow Serving and the SCF service to deliver online predictions.

In practice, the deployed model is applied to two tasks: (1) text‑quality evaluation, where perplexity serves as a quality score that is combined with manual ratings to influence post ranking, and (2) text generation, where a seed word and top‑K sampling produce diverse continuations to assist search query correction and suggestion.

Overall, LSTM‑based language models remain effective for Chinese text applications, and ongoing advances are expected to further improve performance in quality assessment and generation tasks.