Spaces:
Runtime error
Runtime error
domenicrosati
commited on
Commit
Β·
e996282
1
Parent(s):
f2eab41
add treshold for predictions
Browse files
app.py
CHANGED
@@ -11,6 +11,8 @@ import numpy as np
|
|
11 |
from typing import List, Tuple
|
12 |
import torch
|
13 |
|
|
|
|
|
14 |
class CrossEncoder:
|
15 |
def __init__(self, model_path: str, **kwargs):
|
16 |
self.model = CE(model_path, **kwargs)
|
@@ -22,18 +24,21 @@ class CrossEncoder:
|
|
22 |
show_progress_bar=show_progress_bar)
|
23 |
|
24 |
|
25 |
-
SCITE_API_KEY = st.secrets["SCITE_API_KEY"]
|
26 |
-
|
27 |
-
|
28 |
def remove_html(x):
|
29 |
soup = BeautifulSoup(x, 'html.parser')
|
30 |
text = soup.get_text()
|
31 |
return text
|
32 |
|
33 |
|
|
|
|
|
|
|
|
|
|
|
34 |
def search(term, limit=10, clean=True, strict=True):
|
35 |
term = clean_query(term, clean=clean, strict=strict)
|
36 |
# heuristic, 2 searches strict and not? and then merge?
|
|
|
37 |
search = f"https://api.scite.ai/search?mode=citations&term={term}&limit={limit}&offset=0&user_slug=domenic-rosati-keW5&compute_aggregations=false"
|
38 |
req = requests.get(
|
39 |
search,
|
@@ -67,6 +72,7 @@ def find_source(text, docs):
|
|
67 |
'source_title': doc[2],
|
68 |
'source_link': f"https://scite.ai/reports/{doc[0]}"
|
69 |
}
|
|
|
70 |
return None
|
71 |
|
72 |
|
@@ -79,7 +85,7 @@ def init_models():
|
|
79 |
"question-answering", model='sultan/BioM-ELECTRA-Large-SQuAD2-BioASQ8B',
|
80 |
device=device
|
81 |
)
|
82 |
-
reranker = CrossEncoder('cross-encoder/ms-marco-MiniLM-L-
|
83 |
queryexp_tokenizer = AutoTokenizer.from_pretrained("doc2query/all-with_prefix-t5-base-v1")
|
84 |
queryexp_model = AutoModelWithLMHead.from_pretrained("doc2query/all-with_prefix-t5-base-v1")
|
85 |
return question_answerer, reranker, stop, device, queryexp_model, queryexp_tokenizer
|
@@ -98,7 +104,6 @@ def clean_query(query, strict=True, clean=True):
|
|
98 |
return query
|
99 |
|
100 |
|
101 |
-
|
102 |
def card(title, context, score, link, supporting):
|
103 |
st.markdown(f"""
|
104 |
<div class="container-fluid">
|
@@ -138,7 +143,7 @@ st.write("""
|
|
138 |
Ask a scientific question and get an answer drawn from [scite.ai](https://scite.ai) corpus of over 1.1bn citation statements.
|
139 |
Answers are linked to source documents containing citations where users can explore further evidence from scientific literature for the answer.
|
140 |
|
141 |
-
For example try:
|
142 |
""")
|
143 |
|
144 |
st.markdown("""
|
@@ -146,26 +151,27 @@ st.markdown("""
|
|
146 |
""", unsafe_allow_html=True)
|
147 |
|
148 |
with st.expander("Settings (strictness, context limit, top hits)"):
|
|
|
149 |
strict_mode = st.radio(
|
150 |
"Query mode? Strict means all words must match in source snippet. Lenient means only some words must match.",
|
151 |
('lenient', 'strict'))
|
152 |
use_reranking = st.radio(
|
153 |
"Use Reranking? Reranking will rerank the top hits using semantic similarity of document and query.",
|
154 |
('yes', 'no'))
|
|
|
|
|
155 |
use_query_exp = st.radio(
|
156 |
"(Experimental) use query expansion? Right now it just recommends queries",
|
157 |
('yes', 'no'))
|
158 |
-
|
159 |
-
context_lim = st.slider('Context limit? How many documents to use for answering from. Larger is slower but higher quality', 10, 300, 25 if torch.cuda.is_available() else 10)
|
160 |
|
161 |
def paraphrase(text, max_length=128):
|
162 |
input_ids = queryexp_tokenizer.encode(text, return_tensors="pt", add_special_tokens=True)
|
163 |
-
generated_ids = queryexp_model.generate(input_ids=input_ids, num_return_sequences=5, num_beams=5, max_length=max_length)
|
164 |
queries = set([queryexp_tokenizer.decode(g, skip_special_tokens=True, clean_up_tokenization_spaces=True) for g in generated_ids])
|
165 |
preds = '\n * '.join(queries)
|
166 |
return preds
|
167 |
|
168 |
-
|
169 |
def run_query(query):
|
170 |
if use_query_exp == 'yes':
|
171 |
query_exp = paraphrase(f"question2question: {query}")
|
@@ -186,7 +192,6 @@ If you are not getting good results try one of:
|
|
186 |
</div>
|
187 |
</div>
|
188 |
""", unsafe_allow_html=True)
|
189 |
-
|
190 |
if use_reranking == 'yes':
|
191 |
sentence_pairs = [[query, context] for context in contexts]
|
192 |
scores = reranker.predict(sentence_pairs, batch_size=limit, show_progress_bar=False)
|
@@ -195,7 +200,6 @@ If you are not getting good results try one of:
|
|
195 |
context = '\n'.join(sorted_contexts[:context_limit])
|
196 |
else:
|
197 |
context = '\n'.join(contexts[:context_limit])
|
198 |
-
|
199 |
results = []
|
200 |
model_results = qa_model(question=query, context=context, top_k=10)
|
201 |
for result in model_results:
|
@@ -210,14 +214,23 @@ If you are not getting good results try one of:
|
|
210 |
"score": result['score'],
|
211 |
"doi": support["supporting"]
|
212 |
})
|
213 |
-
|
214 |
-
sorted_result = sorted(results, key=lambda x: x['score'], reverse=True)
|
215 |
sorted_result = list({
|
216 |
result['context']: result for result in sorted_result
|
217 |
}.values())
|
218 |
sorted_result = sorted(
|
219 |
sorted_result, key=lambda x: x['score'], reverse=True)
|
220 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
221 |
for r in sorted_result:
|
222 |
answer = r["answer"]
|
223 |
ctx = remove_html(r["context"]).replace(answer, f"<mark>{answer}</mark>").replace(
|
@@ -227,7 +240,6 @@ If you are not getting good results try one of:
|
|
227 |
card(title, ctx, score, r['link'], r['doi'])
|
228 |
|
229 |
query = st.text_input("Ask scientific literature a question", "")
|
230 |
-
|
231 |
if query != "":
|
232 |
with st.spinner('Loading...'):
|
233 |
run_query(query)
|
|
|
11 |
from typing import List, Tuple
|
12 |
import torch
|
13 |
|
14 |
+
SCITE_API_KEY = st.secrets["SCITE_API_KEY"]
|
15 |
+
|
16 |
class CrossEncoder:
|
17 |
def __init__(self, model_path: str, **kwargs):
|
18 |
self.model = CE(model_path, **kwargs)
|
|
|
24 |
show_progress_bar=show_progress_bar)
|
25 |
|
26 |
|
|
|
|
|
|
|
27 |
def remove_html(x):
|
28 |
soup = BeautifulSoup(x, 'html.parser')
|
29 |
text = soup.get_text()
|
30 |
return text
|
31 |
|
32 |
|
33 |
+
# 4 searches: strict y/n, supported y/n
|
34 |
+
# deduplicate
|
35 |
+
# search per query
|
36 |
+
|
37 |
+
|
38 |
def search(term, limit=10, clean=True, strict=True):
|
39 |
term = clean_query(term, clean=clean, strict=strict)
|
40 |
# heuristic, 2 searches strict and not? and then merge?
|
41 |
+
# https://api.scite.ai/search?mode=citations&term=unit%20testing%20software&limit=10&date_from=2000&date_to=2022&offset=0&supporting_from=1&contrasting_from=0&contrasting_to=0&user_slug=domenic-rosati-keW5&compute_aggregations=true
|
42 |
search = f"https://api.scite.ai/search?mode=citations&term={term}&limit={limit}&offset=0&user_slug=domenic-rosati-keW5&compute_aggregations=false"
|
43 |
req = requests.get(
|
44 |
search,
|
|
|
72 |
'source_title': doc[2],
|
73 |
'source_link': f"https://scite.ai/reports/{doc[0]}"
|
74 |
}
|
75 |
+
print("None found for", text)
|
76 |
return None
|
77 |
|
78 |
|
|
|
85 |
"question-answering", model='sultan/BioM-ELECTRA-Large-SQuAD2-BioASQ8B',
|
86 |
device=device
|
87 |
)
|
88 |
+
reranker = CrossEncoder('cross-encoder/ms-marco-MiniLM-L-6-v2', device=device)
|
89 |
queryexp_tokenizer = AutoTokenizer.from_pretrained("doc2query/all-with_prefix-t5-base-v1")
|
90 |
queryexp_model = AutoModelWithLMHead.from_pretrained("doc2query/all-with_prefix-t5-base-v1")
|
91 |
return question_answerer, reranker, stop, device, queryexp_model, queryexp_tokenizer
|
|
|
104 |
return query
|
105 |
|
106 |
|
|
|
107 |
def card(title, context, score, link, supporting):
|
108 |
st.markdown(f"""
|
109 |
<div class="container-fluid">
|
|
|
143 |
Ask a scientific question and get an answer drawn from [scite.ai](https://scite.ai) corpus of over 1.1bn citation statements.
|
144 |
Answers are linked to source documents containing citations where users can explore further evidence from scientific literature for the answer.
|
145 |
|
146 |
+
For example try: Do tanning beds cause cancer?
|
147 |
""")
|
148 |
|
149 |
st.markdown("""
|
|
|
151 |
""", unsafe_allow_html=True)
|
152 |
|
153 |
with st.expander("Settings (strictness, context limit, top hits)"):
|
154 |
+
confidence_threshold = st.slider('Confidence threshold for answering questions? This number represents how confident the model should be in the answers it gives. The number is out of 100%', 0, 100, 1)
|
155 |
strict_mode = st.radio(
|
156 |
"Query mode? Strict means all words must match in source snippet. Lenient means only some words must match.",
|
157 |
('lenient', 'strict'))
|
158 |
use_reranking = st.radio(
|
159 |
"Use Reranking? Reranking will rerank the top hits using semantic similarity of document and query.",
|
160 |
('yes', 'no'))
|
161 |
+
top_hits_limit = st.slider('Top hits? How many documents to use for reranking. Larger is slower but higher quality', 10, 300, 200 if torch.cuda.is_available() else 50)
|
162 |
+
context_lim = st.slider('Context limit? How many documents to use for answering from. Larger is slower but higher quality', 10, 300, 25 if torch.cuda.is_available() else 10)
|
163 |
use_query_exp = st.radio(
|
164 |
"(Experimental) use query expansion? Right now it just recommends queries",
|
165 |
('yes', 'no'))
|
166 |
+
suggested_queries = st.slider('Number of suggested queries to use', 0, 10, 5)
|
|
|
167 |
|
168 |
def paraphrase(text, max_length=128):
|
169 |
input_ids = queryexp_tokenizer.encode(text, return_tensors="pt", add_special_tokens=True)
|
170 |
+
generated_ids = queryexp_model.generate(input_ids=input_ids, num_return_sequences=suggested_queries or 5, num_beams=suggested_queries or 5, max_length=max_length)
|
171 |
queries = set([queryexp_tokenizer.decode(g, skip_special_tokens=True, clean_up_tokenization_spaces=True) for g in generated_ids])
|
172 |
preds = '\n * '.join(queries)
|
173 |
return preds
|
174 |
|
|
|
175 |
def run_query(query):
|
176 |
if use_query_exp == 'yes':
|
177 |
query_exp = paraphrase(f"question2question: {query}")
|
|
|
192 |
</div>
|
193 |
</div>
|
194 |
""", unsafe_allow_html=True)
|
|
|
195 |
if use_reranking == 'yes':
|
196 |
sentence_pairs = [[query, context] for context in contexts]
|
197 |
scores = reranker.predict(sentence_pairs, batch_size=limit, show_progress_bar=False)
|
|
|
200 |
context = '\n'.join(sorted_contexts[:context_limit])
|
201 |
else:
|
202 |
context = '\n'.join(contexts[:context_limit])
|
|
|
203 |
results = []
|
204 |
model_results = qa_model(question=query, context=context, top_k=10)
|
205 |
for result in model_results:
|
|
|
214 |
"score": result['score'],
|
215 |
"doi": support["supporting"]
|
216 |
})
|
217 |
+
sorted_result = sorted(results, key=lambda x: x['score'])
|
|
|
218 |
sorted_result = list({
|
219 |
result['context']: result for result in sorted_result
|
220 |
}.values())
|
221 |
sorted_result = sorted(
|
222 |
sorted_result, key=lambda x: x['score'], reverse=True)
|
223 |
|
224 |
+
if confidence_threshold == 0:
|
225 |
+
threshold = 0
|
226 |
+
else:
|
227 |
+
threshold = (confidence_threshold or 10) / 100
|
228 |
+
|
229 |
+
sorted_result = filter(
|
230 |
+
lambda x: x['score'] > threshold,
|
231 |
+
sorted_result
|
232 |
+
)
|
233 |
+
|
234 |
for r in sorted_result:
|
235 |
answer = r["answer"]
|
236 |
ctx = remove_html(r["context"]).replace(answer, f"<mark>{answer}</mark>").replace(
|
|
|
240 |
card(title, ctx, score, r['link'], r['doi'])
|
241 |
|
242 |
query = st.text_input("Ask scientific literature a question", "")
|
|
|
243 |
if query != "":
|
244 |
with st.spinner('Loading...'):
|
245 |
run_query(query)
|