[MLS-C01] [Algorithms] Text Analysis Algorithms

Defined

Both supervised and unsupervised learning algorithms
Take text or documents as input and either categorize, sequence, or classify the text or documents
Used as preprocessing for many downstream natural language processing (NLP) tasks, such as sentiment analysis, named entity recognition, machine translation
Text classification used for applications that perform web searches, ranking, and document classification

Implements the Word2vec and text classification algorithms
Can use pre-trained vector representations that improve the generalizability of other models that are later trained on a more limited amount of data
Can easily scale for large text datasets
Can train a model on more that a billion words very quickly, in minutes, using a large multi-core CPU or a GPU
Words that are semantically similar correspond to vectors that are close together, resulting that word embeddings capture the semantic relationships between words.
Example use case: market research using sentiment analysis
Important Hyperparameters
1. mode: architecture used for training

Unsupervised learning algorithm that organizes a set of text observations into distinct categories
Frequently used to discover a number of topics shared across documents within a collection of texts, or a corpus
In an LDA algorithm based model, each observation is a document and each feature is a count of a word in the documents
Topics are not specified in advance
Each document is described as a mixture of topics
Example use case: find common topics in call center transcripts
Important Hyperparameters
1. num_topics: number of topics to find in the data
2. feature_dim: size of the vocabulary of the input document corpus
3. min_batch_size: total number of documents in the input document corpus

Unsupervised learning algorithm that organizes a corpus of documents into topics containing word groupings, based on the statistical distribution of the word groupings
Frequently used to classify or summarize documents based on topics detected
Also used to retrieve information or recommend content based on topic similarities
Topics are inferred from observed word distributions in the corpus
Used to visualize the contents of a large set of documents in terms of the learned topics
Similar to LDA, but will produce different outcomes
Example use case: find the topics of newsgroup message posts
Important Hyperparameters
1. feature_dim: vocabulary size of the dataset
2. num_topics: number of required topics

General purpose neural embedding algorithm that finds related clusters of words (words that are semantically similar)
Embeddings can be used to find nearest neighbors of objects, and can also visualize clusters of related objects
Besides word embeddings, Object2Vec can also learn the embeddings of other objects such as sentences, customers, products, etc.
Frequently used for information retrieval, product search, item matching, customer profiling, etc. based on related topics
Supports embeddings of paired tokens, paired sequences, and paired token to sequence
Example use case: recommendation engine based on collaborative filtering
Important Hyperparameters
1. enc0_max_seq_len: maximum sequence length for the enc0 encoder
2. enc0_vocab_size: vocabulary size of enc0 tokens

Supervised learning algorithm with input of a sequence of tokens (audio, text, radar data) and output of another sequence of tokens
Can be used for translation from one language to another, text summarization, speech-to-text
Uses Recurrent Neural Networks (RNNs) and Convolutional Neural Network (CNN) models
Uses state-of-the-art encoder-decoder architecture
Uses input of sequence data in recordio-protobuf format and JSON vocabulary mapping files
Example use case: word pronunciation dictionary, a sequence of text as input and a sequence of audio as output
Important Hyperparameters
1. Has no required hyperparameters

本文由 Oscaner 创作, 采用知识共享署名4.0 国际许可协议进行许可
本站文章除注明转载/出处外, 均为本站原创或翻译, 转载前请务必署名