Initial commit

app.py

@@ -0,0 +1,96 @@
+import streamlit as st
+import pandas as pd
+
+st.set_page_config(layout="wide")
+
+st.sidebar.markdown("**Data Filter Options**")
+split = st.sidebar.selectbox("Dataset Split", ["train", "test"])
+display = st.sidebar.selectbox("Source", ["All", "Only MRPC", "Only MRPC-R1"])
+
+ptype = st.sidebar.radio("Paraphrase Pair Types", ["All",
+                                      "Only Paraphrases (MRPC-R1)", 
+                                      "Only Paraphrases (MRPC)", 
+                                      "Rejected Paraphrases from MRPC",
+                                      "Corrected Paraphrases from MRPC"])
+
+st.sidebar.markdown("**Score Filter Options**")
+filter_by = st.sidebar.selectbox("Filter By Scores From", ["MRPC", "MRPC-R1"])
+display_range_wpd = st.sidebar.slider("Filter by WPD Scores", min_value=0.0, max_value=1.0, value=(0.1,0.7))
+display_range_ld = st.sidebar.slider("Filter by LD Scores", min_value=0.0, max_value=1.0, value=(0.1,0.4))
+
+st.sidebar.markdown("""**Explanation**
+This demo allows you to explore the data inside [MRPC](https://www.microsoft.com/en-us/download/details.aspx?id=52398), showing how we can use Word Position Deviation (WPD) and Lexical Deviation (LD) to find different types of paraphrases. By using what we observe from the data, we can also correct numerous labelling errors inside MRPC, presenting the a revision of MRPC termed as MRPC-R1. This demo accompanies the paper ["Towards Better Characterization of Paraphrases" (ACL 2022)](https://github.com/tlkh/paraphrase-metrics).""")
+
+st.markdown("**MRPC Paraphrase Data Explorer**")
+
+def load_df(split):
+    if split == "train":
+        df = pd.read_csv("./mrpc_train_scores.csv")
+    else:
+        df = pd.read_csv("./mrpc_test_scores.csv")
+    df.reset_index(drop=True, inplace=True)
+    return df
+
+def filter_df(df, display, ptype, filter_by):
+    # filter data
+    if display == "MRPC":
+        df = df.drop(["new_s1", "new_s2"], axis=1)
+    elif display == "MRPC-R1":
+        df = df.drop(["og_s1", "og_s2"], axis=1)
+    # filter paraphrase type
+    if ptype == "Only Paraphrases (MRPC)":
+        condition = df.og_label == 1
+        df_sel = df[condition]
+    elif ptype == "Only Paraphrases (MRPC-R1)":
+        condition = df.new_label == 1
+        df_sel = df[condition]
+    elif ptype == "Rejected Paraphrases from MRPC":
+        condition = (df.new_label == 0) & (df.og_label == 1)
+        df_sel = df[condition]
+    elif ptype == "Corrected Paraphrases from MRPC":
+        condition = df.remarks == "corrected"
+        df_sel = df[condition]
+    else:
+        # all
+        df_sel = df
+    # sort by scores
+    if filter_by == "MRPC":
+        # wpd
+        condition = (df_sel.og_wpd >= display_range_wpd[0]) & (df_sel.og_wpd < display_range_wpd[1])
+        df_sel = df_sel[condition]
+        # ld
+        condition = (df_sel.og_ld >= display_range_ld[0]) & (df_sel.og_ld < display_range_ld[1])
+        df_sel = df_sel[condition]
+    else:
+        # wpd
+        condition = (df_sel.new_wpd >= display_range_wpd[0]) & (df_sel.new_wpd < display_range_wpd[1])
+        df_sel = df_sel[condition]
+        # ld
+        condition = (df_sel.new_ld >= display_range_ld[0]) & (df_sel.new_ld < display_range_ld[1])
+        df_sel = df_sel[condition]
+    # filter scores
+    if filter_by == "MRPC":
+        df_sel.sort_values("og_ld", inplace=True)
+        df_sel.sort_values("og_wpd", inplace=True)
+    else:
+        df_sel.sort_values("new_ld", inplace=True)
+        df_sel.sort_values("new_wpd", inplace=True)
+    return df_sel
+    
+
+df = load_df(split)
+
+df_sel = filter_df(df, display, ptype, filter_by)
+
+# CSS to inject contained in a string
+hide_table_row_index = """
+            <style>
+            tbody th {display:none}
+            .blank {display:none}
+            </style>
+            """
+
+# inject css at the end
+st.markdown("Total "+str(len(df_sel))+" items"+hide_table_row_index, unsafe_allow_html=True)
+
+st.table(data=df_sel)

requirements.txt

@@ -0,0 +1,2 @@
+pandas
+streamlit