elisno

elisno

elisno

# Plot two figures
plt.subplots(nrows=1, ncols=2, figsize=(20, 10))


# LEFT PLOT
# Plot scatter plot of umap embeddings with clusterer labels as colors

x_plot, y_plot = selected_outlier_subset_embeddings_umap[:, 0], selected_outlier_subset_embeddings_umap[:, 1]
plt.subplot(1, 2, 1)
for i, topic in enumerate(np.unique(cluster_labels)):

elisno

n = 7

top_idx = np.argpartition(tf_idf, -n)[:, -n:]
vlist = list(vocab)
for c, _class in enumerate(classes.keys()):
    topn_idx = top_idx[c, :]
    topn_terms = [vlist[idx] for idx in topn_idx]
    if _class != -1:
        print(f"Topic class {_class}: {topn_terms}")
    else:

elisno

###### Create documents from sentence pairs

# Get combined text from the selected outliers
# Joining the premise and hypothesis together
def join_sentence_pair(example):
    docs = []
    for premise, hypothesis in zip(example["premise"], example["hypothesis"]):
        docs.append(premise + " " + hypothesis)
    example["docs"] = docs
    return example

elisno

clusterer = hdbscan.HDBSCAN(min_cluster_size=6, min_samples=4)
clusterer.fit(selected_outlier_subset_embeddings_umap)
cluster_labels = clusterer.labels_

clusterer.condensed_tree_.plot(select_clusters=True)

# plot each set of points in a different color
plt.figure(figsize=(10, 10))
for i in np.unique(cluster_labels):
    if i != -1:

elisno

# Get embeddings of selected outliers
selected_outlier_subset_embeddings = test_feature_embeddings[outlier_ids]

# Reduce dimensionality with UMAP
umap_fit = UMAP(n_components=2, n_neighbors=8, random_state=SEED)
selected_outlier_subset_embeddings_umap = umap_fit.fit_transform(selected_outlier_subset_embeddings)

# Set plot labels
mismatched_labels = {"nineeleven": 0, "facetoface": 1, "letters": 2, "oup": 3, "verbatim": 4}
matched_labels = {"fiction": 5, "government": 6, "slate": 7, "telephone": 8, "travel": 9}

elisno

# Get embeddings of test examples whose outlier scores are below the threshold

sorted_ids = test_outlier_scores.argsort()
outlier_scores = test_outlier_scores[sorted_ids]
outlier_ids = sorted_ids[outlier_scores < threshold]

selected_outlier_subset = test_data.select(outlier_ids)
selected_outlier_subset.to_pandas().tail(15)

elisno

# Take the 2.5th percentile of the outlier scores in the training data as the threshold
threshold = np.percentile(test_outlier_scores, 2.5)

fig, axes = plt.subplots(nrows=1, ncols=2, figsize=(10, 5))
plt_range = [min(train_outlier_scores.min(),test_outlier_scores.min()), \
             max(train_outlier_scores.max(),test_outlier_scores.max())]

axes[0].hist(train_outlier_scores, range=plt_range, bins=50)
axes[0].set(title='train_outlier_scores distribution', ylabel='Frequency')
axes[0].axvline(x=threshold, color='red', linewidth=2)

elisno

# Visualize 15 least severe outliers in test data
bottom_outlier_idxs = (-test_outlier_scores).argsort()[:20]
bottom_outlier_subset = test_data.select(bottom_outlier_idxs)
bottom_outlier_subset.to_pandas()