问题

1. 使用预定义的方面来生成情感分数

import torch
from transformers import AutoTokenizer, AutoModelForSequenceClassification
import pandas as pd

# 加载ABSA模型和分词器
model_name = "yangheng/deberta-v3-base-absa-v1.1"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForSequenceClassification.from_pretrained(model_name)

# 生成方面和情感
aspects = []
sentiments = []

for index, row in df.iterrows():
    text = row['text']
    row_aspects = row['aspects']
    
    aspect_sentiments = []
    
    for aspect in row_aspects:
        inputs = tokenizer(text, aspect, return_tensors="pt")
        
        with torch.inference_mode():
            outputs = model(**inputs)
        
        predicted_sentiment = torch.argmax(outputs.logits).item()
        sentiment_label = model.config.id2label[predicted_sentiment]
        
        aspect_sentiments.append(f"{aspect}: {sentiment_label}")
    
    aspects.append(row_aspects)
    sentiments.append(aspect_sentiments)

# 将生成的方面和情感添加到DataFrame中
df['generated_aspects'] = aspects
df['generated_sentiments'] = sentiments

# 打印更新后的DataFrame
print(df)

2. 生成文本的方面和相应的情感分数

import torch
import torch.nn.functional as F
from transformers import AutoTokenizer, AutoModelForSequenceClassification

model_name = "yangheng/deberta-v3-base-absa-v1.1"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForSequenceClassification.from_pretrained(model_name)

aspects = ["food", "service"]
text = "The food was great but the service was terrible."
sentiment_aspect = {}

for aspect in aspects:
    inputs = tokenizer(text, aspect, return_tensors="pt")

    with torch.inference_mode():
        outputs = model(**inputs)

    scores = F.softmax(outputs.logits[0], dim=-1)
    label_id = torch.argmax(scores).item()
    sentiment_aspect[aspect] = (model.config.id2label[label_id], scores[label_id].item())

print(sentiment_aspect)

data = {
    'text': [
        "The camera quality of this phone is amazing.",
        "The belt is poor quality",
        "The battery life could be improved.",
        "The display is sharp and vibrant.",
        "The customer service was disappointing."
    ],
    'aspects': [
        ["camera", "phone"],
        [],
        ["battery", "life"],
        ["display"],
        ["customer service"]
    ]
}

df = pd.DataFrame(data)

import torch
from transformers import AutoTokenizer, AutoModelForSequenceClassification
import pandas as pd

# Load the ABSA model and tokenizer
model_name = "yangheng/deberta-v3-base-absa-v1.1"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForSequenceClassification.from_pretrained(model_name)




# Generate aspects and sentiments
aspects = []
sentiments = []

for index, row in df.iterrows():
    text = row['text']
    row_aspects = row['aspects']
    
    aspect_sentiments = []
    
    for aspect in row_aspects:
        inputs = tokenizer(text, aspect, return_tensors="pt")
        
        with torch.inference_mode():
            outputs = model(**inputs)
        
        predicted_sentiment = torch.argmax(outputs.logits).item()
        sentiment_label = model.config.id2label[predicted_sentiment]
        
        aspect_sentiments.append(f"{aspect}: {sentiment_label}")
    
    aspects.append(row_aspects)
    sentiments.append(aspect_sentiments)

# Add the generated aspects and sentiments to the DataFrame
df['generated_aspects'] = aspects
df['generated_sentiments'] = sentiments

# Print the updated DataFrame
print(df)

import torch
import torch.nn.functional as F
from transformers import AutoTokenizer, AutoModelForSequenceClassification

model_name = "yangheng/deberta-v3-base-absa-v1.1"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForSequenceClassification.from_pretrained(model_name)

aspects = ["food", "service"]
text = "The food was great but the service was terrible."
sentiment_aspect = {}
for aspect in aspects:
  inputs = tokenizer(text, aspect, return_tensors="pt")

  with torch.inference_mode():
    outputs = model(**inputs)

  scores = F.softmax(outputs.logits[0], dim=-1)
  label_id = torch.argmax(scores).item()
  sentiment_aspect[aspect] = (model.config.id2label[label_id], scores[label_id].item())

print(sentiment_aspect)

答案1

得分: 2

# Load the ABSA model and tokenizer
model_name = "yangheng/deberta-v3-base-absa-v1.1"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForSequenceClassification.from_pretrained(model_name)

classifier = pipeline("text-classification", model=model, tokenizer=tokenizer)

for aspect in ['camera', 'phone']:
   print(aspect, classifier('The camera quality of this phone is amazing.',  text_pair=aspect))

[out]:

camera [{'label': 'Positive', 'score': 0.9967294931411743}]
phone [{'label': 'Neutral', 'score': 0.9472787380218506}]

To get the zero-shot classification scores in general, try using pipeline:

from transformers import AutoTokenizer, AutoModelForSequenceClassification
from transformers import pipeline

# Load the ABSA model and tokenizer
model_name = "yangheng/deberta-v3-base-absa-v1.1"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForSequenceClassification.from_pretrained(model_name)

pipe = pipeline("zero-shot-classification", model=model, tokenizer=tokenizer)

pipe("The camera quality of this phone is amazing.", candidate_labels=["camera", "phone"])

[out]:

{'sequence': 'The camera quality of this phone is amazing.',
 'labels': ['camera', 'phone'],
 'scores': [0.9036691784858704, 0.09633082151412964]}

Depending on what "text generated aspect" means, perhaps it's keyword extraction, and if so, doing a search on https://huggingface.co/models?search=keyword, gives this as the top downloaded model, https://huggingface.co/yanekyuk/bert-uncased-keyword-extractor

from transformers import AutoTokenizer, AutoModelForTokenClassification

tokenizer2 = AutoTokenizer.from_pretrained("yanekyuk/bert-uncased-keyword-extractor")
model2 = AutoModelForTokenClassification.from_pretrained("yanekyuk/bert-uncased-keyword-extractor")

def extract_aspect(text):
    extractor = pipeline("ner", model=model2, tokenizer=tokenizer2)
    phrasesids = []
    for tag in extractor(text):
        if tag['entity'].startswith('B'):
            phrasesids.append([tag['start'], tag['end']])
        if tag['entity'].startswith('I'):
            phrasesids[-1][-1] = tag['end']
    phrases = [text
:p[1]] for p in phrasesids]
    return phrases

text = "The camera quality of this phone is amazing."

extract_aspect(text)

[out]:

camera

Putting the extractor and classifier together:

from transformers import AutoTokenizer, AutoModelForSequenceClassification, AutoModelForTokenClassification
from transformers import pipeline

# Load the ABSA model and tokenizer
model_name = "yangheng/deberta-v3-base-absa-v1.1"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForSequenceClassification.from_pretrained(model_name)

classifier = pipeline("zero-shot-classification", model=model, tokenizer=tokenizer)

tokenizer2 = AutoTokenizer.from_pretrained("yanekyuk/bert-uncased-keyword-extractor")
model2 = AutoModelForTokenClassification.from_pretrained("yanekyuk/bert-uncased-keyword-extractor")

def extract_aspect(text):
    extractor = pipeline("ner", model=model2, tokenizer=tokenizer2)
    phrasesids = []
    for tag in extractor(text):
        if tag['entity'].startswith('B'):
            phrasesids.append([tag['start'], tag['end']])
        if tag['entity'].startswith('I'):
            phrasesids[-1][-1] = tag['end']
    phrases = [text
:p[1]] for p in phrasesids]
    return phrases

text = "The camera quality of this phone is amazing."

pipe(text, candidate_labels=extract_aspect(text))

[out]:

{'sequence': 'The camera quality of this phone is amazing.',
 'labels': ['camera'],
 'scores': [0.9983300566673279]}

Q: But the extracted keywords is not "right" or doesn't match the pre-defined ones?

A: No model is perfect and the model example above is a keyword extractor not a product aspect extractor. YMMV.

Q: Why isn't the zero-shot classifier giving me negative / positive labels?

A: The zero-shot classifier is labeling the data based on the extracted labels. Not a sentiment classifier.

<details>
<summary>英文:</summary>

Specific to the `yangheng/deberta-v3-base-absa-v1.1` model this is the usage and you have to loop through the model one time per aspect:

Load the ABSA model and tokenizer

model_name = "yangheng/deberta-v3-base-absa-v1.1"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForSequenceClassification.from_pretrained(model_name)

classifier = pipeline("text-classification", model=model, tokenizer=tokenizer)

for aspect in ['camera', 'phone']:
print(aspect, classifier('The camera quality of this phone is amazing.', text_pair=aspect))

[out]:

camera [{'label': 'Positive', 'score': 0.9967294931411743}]
phone [{'label': 'Neutral', 'score': 0.9472787380218506}]

----

To get the zero-shot classification scores in general, try using `pipeline`:

from transformers import AutoTokenizer, AutoModelForSequenceClassification
from transformers import pipeline

Load the ABSA model and tokenizer

model_name = "yangheng/deberta-v3-base-absa-v1.1"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForSequenceClassification.from_pretrained(model_name)

pipe = pipeline("zero-shot-classification", model=model, tokenizer=tokenizer)

pipe("The camera quality of this phone is amazing.", candidate_labels=["camera", "phone"])

[out]:

{'sequence': 'The camera quality of this phone is amazing.',
'labels': ['camera', 'phone'],
'scores': [0.9036691784858704, 0.09633082151412964]}

---- 

Depending on what "text generated aspect" means, perhaps it's keyword extraction, and if so, doing a search on https://huggingface.co/models?search=keyword, gives this as the top downloaded model, https://huggingface.co/yanekyuk/bert-uncased-keyword-extractor

from transformers import AutoTokenizer, AutoModelForTokenClassification

tokenizer2 = AutoTokenizer.from_pretrained("yanekyuk/bert-uncased-keyword-extractor")
model2 = AutoModelForTokenClassification.from_pretrained("yanekyuk/bert-uncased-keyword-extractor")

def extract_aspect(text):
extractor = pipeline("ner", model=model2, tokenizer=tokenizer2)
phrasesids = []
for tag in extractor(text):
if tag['entity'].startswith('B'):
phrasesids.append([tag['start'], tag['end']])
if tag['entity'].startswith('I'):
phrasesids[-1][-1] = tag['end']
phrases = [text

:p1] for p in phrasesids]
return phrases

text = "The camera quality of this phone is amazing."

extract_aspect(text)

[out]:

camera

Putting the extractor and classifier together:

from transformers import AutoTokenizer, AutoModelForSequenceClassification, AutoModelForTokenClassification
from transformers import pipeline

Load the ABSA model and tokenizer

model_name = "yangheng/deberta-v3-base-absa-v1.1"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForSequenceClassification.from_pretrained(model_name)

classifier = pipeline("zero-shot-classification", model=model, tokenizer=tokenizer)

tokenizer2 = AutoTokenizer.from_pretrained("yanekyuk/bert-uncased-keyword-extractor")
model2 = AutoModelForTokenClassification.from_pretrained("yanekyuk/bert-uncased-keyword-extractor")

def extract_aspect(text):
extractor = pipeline("ner", model=model2, tokenizer=tokenizer2)
phrasesids = []
for tag in extractor(text):
if tag['entity'].startswith('B'):
phrasesids.append([tag['start'], tag['end']])
if tag['entity'].startswith('I'):
phrasesids[-1][-1] = tag['end']
phrases = [text

:p1] for p in phrasesids]
return phrases

text = "The camera quality of this phone is amazing."

pipe(text, candidate_labels=extract_aspect(text))

[out]:

{'sequence': 'The camera quality of this phone is amazing.',
'labels': ['camera'],
'scores': [0.9983300566673279]}

-----

### Q: But the extracted keywords is not "right" or doesn't match the pre-defined ones?

A: No model is perfect and the model example above is a keyword extractor not a product aspect extractor. YMMV.

### Q: Why isn't the zero-shot classifier giving me negative / positive labels? 

A: The zero-shot classifier is labelling the data based on the extracted labels. Not a sentiment classifier. 

</details>