Training

This model adapter is designed for token classification tasks, enabling it to extract aspect terms and predict the sentiment polarity associated with the extracted aspect terms. The extracted aspect terms will be the span(s) from the input text on which a sentiment is being expressed. It has been created using PEFT framework for LoRA:Low-Rank Adaptation.

Datasets

This model has been trained on the following datasets:

1. Aspect Based Sentiment Analysis SemEval Shared Tasks (2014, 2015, 2016)
2. Multi-Aspect Multi-Sentiment MAMS

Use

• Loading the base model and merging it with LoRA parameters

from transformers import AutoTokenizer, AutoModelForTokenClassification
from peft import PeftModel

# preparing the labels
labels = {"B-neu": 1, "I-neu": 2, "O": 0, "B-neg": 3, "B-con": 4, "I-pos": 5, "B-pos": 6, "I-con": 7, "I-neg": 8, "X": -100}
id2labels = {k:lab for lab, k in labels.items()}
labels2ids = {k:lab for lab, k in id2labels.items()}

# loading tokenizer and base_model
base_id = 'FacebookAI/roberta-large'
tokenizer = AutoTokenizer.from_pretrained(base_id)
base_model = AutoModelForTokenClassification.from_pretrained(base_id, num_labels=len(labels), id2label=id2labels, label2id=labels2ids)

# using this adapter with base model 
model = PeftModel.from_pretrained(base_model, 'gauneg/roberta-large-absa-ate-sentiment-lora-adapter', is_trainable=False)

This model can be utilized in the following two methods:

1. Using pipelines for end to end inference
2. Making token level inference

Using end-to-end token classification pipeline

# after loading base model and the adapter as shown in the previous snippet
from transformers import pipeline

ate_senti_pipeline = pipeline(task='ner', 
                  aggregation_strategy='simple',
                  model=model,
                  tokenizer=tokenizer)


text_input = "Been here a few times and food has always been good but service really suffers when it gets crowded."
ate_senti_pipeline(text_input)

OUTPUT

[{'entity_group': 'pos',
  'score': 0.92310727,
  'word': ' food',
  'start': 26,
  'end': 30},
 {'entity_group': 'neg',
  'score': 0.90695626,
  'word': ' service',
  'start': 56,
  'end': 63}]

OR

Making token level inference

# after loading base model and the adapter as shown in the previous snippet

# making one prediction at a time (should be padded/batched and truncated for efficiency)
text_input = "Been here a few times and food has always been good but service really suffers when it gets crowded."
tok_inputs = tokenizer(text_input, return_tensors="pt")


y_pred = model(**tok_inputs) # predicting the logits

# since first and the last tokens are excluded (<s> and </s>) labels predicted against them
# should be removed before decoding predicted labels
y_pred_fin = y_pred.logits.argmax(dim=-1)[0][1:-1] # selecting the most favoured labels for each token from the logits


decoded_pred = [id2lab[logx.item()] for logx in y_pred_fin]


## displaying the input tokens with predictions and skipping <s> and </s> tokens at the beginning and the end respectively
decoded_toks = tok_inputs['input_ids'][0][1:-1]
tok_levl_pred = list(zip(tokenizer.convert_ids_to_tokens(decoded_toks), decoded_pred))

RESULTS in tok_levl_pred variable:

[('Be', 'O'),
 ('en', 'O'),
 ('Ġhere', 'O'),
 ('Ġa', 'O'),
 ('Ġfew', 'O'),
 ('Ġtimes', 'O'),
 ('Ġand', 'O'),
 ('Ġfood', 'B-pos'),
 ('Ġhas', 'O'),
 ('Ġalways', 'O'),
 ('Ġbeen', 'O'),
 ('Ġgood', 'O'),
 ('Ġbut', 'O'),
 ('Ġservice', 'B-neg'),
 ('Ġreally', 'O'),
 ('Ġsuffers', 'O'),
 ('Ġwhen', 'O'),
 ('Ġit', 'O'),
 ('Ġgets', 'O'),
 ('Ġcrowded', 'O'),
 ('.', 'O')]

Evaluation on Benchmark Test Datasets

The first evaluation is for token-extraction task without considering the polarity of the extracted tokens. The tokens expected to be extracted are aspect term tokens on which the sentiments have been expressed. (scores are expressed as micro-averages of B-I-O labels)

ATE (Aspect Term Extraction Only)

Aspect Sentiment Evaluation

This evaluation considers token-extraction task with polarity of the extracted tokens. The tokens expected to be extracted are aspect term tokens on which the sentiments have been expressed along with the polarity of the sentiments. (scores are expressed as macro-averages)