diff --git a/src/pubget/_vectorization.py b/src/pubget/_vectorization.py
index b0db48c..418bd1c 100644
--- a/src/pubget/_vectorization.py
+++ b/src/pubget/_vectorization.py
@@ -165,7 +165,6 @@ def _vectorize_articles(
 
     Returns the pmcids and the mapping text field: csr matrix of features.
     """
-    articles.fillna("", inplace=True)
     vectorized = {}
     for field in _FIELDS:
         vectorized[field] = vectorizer.transform(articles[field].values)
@@ -185,7 +184,15 @@ def _extract_word_counts(
     ).fit()
     chunksize = 200
     with open(corpus_file, encoding="utf-8") as corpus_fh:
-        all_chunks = pd.read_csv(corpus_fh, chunksize=chunksize)
+        # The text fields are read as strings and missing values as empty
+        # strings: pandas would otherwise give an empty field the float dtype,
+        # and the vectorizer expects text.
+        all_chunks = pd.read_csv(
+            corpus_fh,
+            chunksize=chunksize,
+            dtype=dict.fromkeys(_FIELDS, str),
+            keep_default_na=False,
+        )
         vectorized_chunks = Parallel(n_jobs=n_jobs, verbose=8)(
             delayed(_vectorize_articles)(chunk, vectorizer=vectorizer)
             for chunk in all_chunks
diff --git a/src/pubget/_vocabulary.py b/src/pubget/_vocabulary.py
index a1dc5a4..ac7ac91 100644
--- a/src/pubget/_vocabulary.py
+++ b/src/pubget/_vocabulary.py
@@ -21,13 +21,21 @@
 _LOG = logging.getLogger(__name__)
 _STEP_NAME = "extract_vocabulary"
 _STEP_DESCRIPTION = "Extract vocabulary of word n-grams from text."
+_TEXT_FIELDS = ("title", "keywords", "abstract", "body")
 
 
 def _iter_corpus(corpus_fh: TextIO) -> Generator[str, None, None]:
     """Yield the concatenated text fields of articles one by one."""
     n_articles = 0
-    for chunk in pd.read_csv(corpus_fh, chunksize=500):
-        chunk.fillna("", inplace=True)
+    # The text fields are read as strings and missing values as empty strings:
+    # pandas would otherwise give an empty field the float dtype, and refuse to
+    # store the empty strings in it.
+    for chunk in pd.read_csv(
+        corpus_fh,
+        chunksize=500,
+        dtype=dict.fromkeys(_TEXT_FIELDS, str),
+        keep_default_na=False,
+    ):
         text = chunk["title"].str.cat(
             chunk.loc[:, ["keywords", "abstract", "body"]], sep="\n"
         )
diff --git a/tests/test_data_extraction.py b/tests/test_data_extraction.py
index 536a2ac..36aa351 100644
--- a/tests/test_data_extraction.py
+++ b/tests/test_data_extraction.py
@@ -54,7 +54,8 @@ def _check_extracted_data(data_dir, articles_with_coords_only):
     assert text.shape == (n_articles, 5)
     assert text.at[0, "body"].strip().startswith("The text")
     coordinates = pd.read_csv(data_dir.joinpath("coordinates.csv"))
-    assert coordinates.shape == (12, 6)
+    # 5 articles with a 2-row coordinate table and one with a 10-row one
+    assert coordinates.shape == (20, 6)
     authors = pd.read_csv(data_dir.joinpath("authors.csv"))
     assert authors.shape == (n_authors, 3)
     assert authors["pmcid"].nunique() == n_articles
@@ -104,9 +105,16 @@ def test_extract_data_to_csv_with_tables(tmp_path, articles_dir):
         articles_dir, tmp_path.joinpath("extracted_data"), keep_tables=True
     )
     assert code == ExitCode.COMPLETED
-    text = pd.read_csv(data_dir.joinpath("text.csv"))
-    assert text.at[0, "body"].strip().startswith("The text")
-    assert "X\tY\tZ\n10\t20\t30\n" in text.at[0, "body"]
+    # rows are in the order in which the article directories are visited, so
+    # articles are looked up by pmcid rather than by position.
+    text = pd.read_csv(data_dir.joinpath("text.csv"), index_col="pmcid")
+    body = text.at[9054084, "body"]
+    assert body.strip().startswith("The text")
+    assert "X\tY\tZ\n10\t20\t30\n" in body
+    # a table with many columns is inserted in full, not truncated
+    wide_table_body = text.at[9056519, "body"]
+    assert "X\tY\tZ\tRegion\tActivation_Level" in wide_table_body
+    assert "-30\t-40\t-50\tBasal Ganglia" in wide_table_body
 
 
 def test_extractor_failures(articles_dir, tmp_path, monkeypatch):
diff --git a/tests/test_labelbuddy.py b/tests/test_labelbuddy.py
index ab89937..2716ad9 100644
--- a/tests/test_labelbuddy.py
+++ b/tests/test_labelbuddy.py
@@ -1,5 +1,6 @@
 import json
 import math
+import re
 from unittest.mock import Mock
 
 import pandas as pd
@@ -65,7 +66,9 @@ def test_make_labelbuddy_documents(
     ) as f:
         docs = [json.loads(doc_json) for doc_json in f]
     assert len(docs) == expected_batch_size
-    assert all("Body\n The text of" in d["text"] for d in docs)
+    # the whitespace between the "Body" heading and the start of the body
+    # depends on how the article's XML is indented
+    assert all(re.search(r"# Body\s+The text of", d["text"]) for d in docs)
     _check_batch_info(labelbuddy_dir)