Tagging Sentiment with Flair Standard Models

David (Yen-Chieh) Liao

Postdoc at Text & Policy Research Group and SPIRe in UCD

An Example Using sentiment Model (Pre-trained English Model)

library(flaiR)
data("uk_immigration")
uk_immigration <- head(uk_immigration, 5)

Download the English sentiment model from FlairNLP on Hugging Face. Currently, it also supports a large English sentiment model and a German pre-trained model.

tagger_sent <- load_tagger_sentiments("sentiment")

Flair NLP operates under the PyTorch framework. As such, we can use the $to method to set the device for the Flair Python library. The flair_device(“cpu”) allows you to select whether to use the CPU, CUDA devices (like cuda:0, cuda:1, cuda:2), or specific MPS devices on Mac (such as mps:0, mps:1, mps:2). For information on Accelerated PyTorch training on Mac, please refer to https://developer.apple.com/metal/pytorch/. For more about CUDA, please visit: https://developer.nvidia.com/cuda-zone

tagger_sent$to(flair_device("mps")) 

TextClassifier(
  (embeddings): TransformerDocumentEmbeddings(
    (model): DistilBertModel(
      (embeddings): Embeddings(
        (word_embeddings): Embedding(30522, 768, padding_idx=0)
        (position_embeddings): Embedding(512, 768)
        (LayerNorm): LayerNorm((768,), eps=1e-12, elementwise_affine=True)
        (dropout): Dropout(p=0.1, inplace=False)
      )
      (transformer): Transformer(
        (layer): ModuleList(
          (0-5): 6 x TransformerBlock(
            (attention): MultiHeadSelfAttention(
              (dropout): Dropout(p=0.1, inplace=False)
              (q_lin): Linear(in_features=768, out_features=768, bias=True)
              (k_lin): Linear(in_features=768, out_features=768, bias=True)
              (v_lin): Linear(in_features=768, out_features=768, bias=True)
              (out_lin): Linear(in_features=768, out_features=768, bias=True)
            )
            (sa_layer_norm): LayerNorm((768,), eps=1e-12, elementwise_affine=True)
            (ffn): FFN(
              (dropout): Dropout(p=0.1, inplace=False)
              (lin1): Linear(in_features=768, out_features=3072, bias=True)
              (lin2): Linear(in_features=3072, out_features=768, bias=True)
              (activation): GELUActivation()
            )
            (output_layer_norm): LayerNorm((768,), eps=1e-12, elementwise_affine=True)
          )
        )
      )
    )
  )
  (decoder): Linear(in_features=768, out_features=2, bias=True)
  (dropout): Dropout(p=0.0, inplace=False)
  (locked_dropout): LockedDropout(p=0.0)
  (word_dropout): WordDropout(p=0.0)
  (loss_function): CrossEntropyLoss()
)

results <- get_sentiments(uk_immigration$text, seq_len(nrow(uk_immigration)),
                          tagger_sent)

print(results)
#>    doc_id sentiment     score
#>     <int>    <char>     <num>
#> 1:      1  POSITIVE 0.8097584
#> 2:      2  POSITIVE 0.9990165
#> 3:      3  POSITIVE 0.8827484
#> 4:      4  NEGATIVE 0.9997155
#> 5:      5  POSITIVE 0.8604351

Batch Processing in English Sentiment Model

Processing texts individually can be both inefficient and memory-intensive. On the other hand, processing all the texts simultaneously could surpass memory constraints, especially if each document in the dataset is sizable. Parsing the documents in smaller batches may provide an optimal compromise between these two scenarios. Batch processing can enhance efficiency and aid in memory management.

By default, the batch_size parameter is set to 5. You can consider starting with this default value and then experimenting with different batch sizes to find the one that works best for your specific use case. You can monitor memory usage and processing time to help you make a decision. If you have access to a GPU, you might also try larger batch sizes to take advantage of GPU parallelism. However, be cautious not to set the batch size too large, as it can lead to out-of-memory errors. Ultimately, the choice of batch size should be based on a balance between memory constraints, processing efficiency, and the specific requirements of your entity extraction task.

batch_process_results  <- get_sentiments_batch(uk_immigration$text,
                                               uk_immigration$speaker, 
                                               tagger_sent, 
                                               show.text_id = FALSE,
                                               batch_size = 2,
                                               verbose = TRUE)
#> CPU is used.
#> Processing batch 1 out of 3...
#> Processing batch 2 out of 3...
#> Processing batch 3 out of 3...

print(batch_process_results)
#>              doc_id sentiment     score
#>              <char>    <char>     <num>
#> 1: Philip Hollobone  POSITIVE 0.8097584
#> 2:  Stewart Jackson  POSITIVE 0.9990165
#> 3: Philip Hollobone  POSITIVE 0.8827485
#> 4:  Stewart Jackson  NEGATIVE 0.9997155
#> 5: Philip Hollobone  POSITIVE 0.8604351