ML-405/ML-427 - OntologyElement improvements (#3758)

- the "value" attribute from <input/> tag will be taken into account and
processed as "text" in ontology
- the tables will now be parsed without any ids and classes - we have
different reasons behind that, for example, embeddings with ids and
classes can lose some semantic value. Also, more tokens = more expensive
LLM call
-  cleaned to_html, created to_text for OntologyElement

Author: MaksOpp

CHANGELOG.md

@@ -1,7 +1,11 @@
-## 0.16.4-dev0
+## 0.16.4-dev1
 
 ### Enhancements
 
+* **`value` attribute in `<input/>` element is parsed to `OntologyElement.text` in ontology**
+* **`id` and `class` attributes removed from Table subtags in HTML partitioning**
+* **cleaned `to_html` and newly introduced `to_text` in `OntologyElement`**
+
 ### Features
 
 ### Fixes

test_unstructured/documents/unstructured_json_output/example.json

@@ -56,7 +56,7 @@
             "parent_id": "3a6b156a81764e17be128264241f8136",
             "text_as_html": "<form class=\"Form\" id=\"637c2f6935fb4353a5f73025ce04619d\"> <label class=\"FormField\" for=\"company-name\" id=\"50027cccbe1948c9853ce0de37b635c2\">From field name </label><input class=\"FormFieldValue\" id=\"0032242af75c4b37984ea7fea9aac74c\" value=\"Example value\" /></form>"
         },
-        "text": "From field name",
+        "text": "From field name Example value",
         "type": "UncategorizedText"
     },
     {
@@ -78,9 +78,9 @@
             "filename": "example.pdf",
             "page_number": 1,
             "parent_id": "592422373ed741b68a077e2003f8ed81",
-            "text_as_html": "<table class=\"Table\" id=\"dc3792d4422e444f90876b56d0cfb20d\"> <thead class=\"TableHeader\" id=\"50a5548a87e84024af590b3d2830d140\"> <tr class=\"TableRow\" id=\"5e473d7742474412be72dc4e2c45bd4a\"> <th class=\"TableCellHeader\" id=\"01800309aa42411c98ae30f85b23f399\">Description </th><th class=\"TableCellHeader\" id=\"c2765b63d08946a2851955e79e301de4\">Row header </th></tr></thead><tbody class=\"TableBody\" id=\"e0a9a8ffdd7148ad8b4a274b073d340a\"> <tr class=\"TableRow\" id=\"77e829974632455191330b0b8545d1e3\"> <td class=\"TableCell\" id=\"7fee12d4c5554b7da778d6f8fdec8a57\">Value description </td><td class=\"TableCell\" id=\"5a7a33b0c57b4eb881a35bce9f87c831\"> <span class=\"Currency\" id=\"87220f9d62c3482e92e7de72a26869cd\">50 $ </span><span class=\"Measurement\" id=\"0095b9efb90a4cca991e73547c7165f1\">(1.32 %) </span></td></tr></tbody></table>"
+            "text_as_html": "<table class=\"Table\" id=\"dc3792d4422e444f90876b56d0cfb20d\"> <thead> <tr> <th>Description</th><th>Row header</th></tr></thead><tbody> <tr> <td>Value description</td><td>50 $ (1.32 %)</td></tr></tbody></table>"
         },
-        "text": "Description Row header   Value description  50 $ (1.32 %)",
+        "text": "Description Row header Value description 50 $ (1.32 %)",
         "type": "Table"
     },
     {

test_unstructured/partition/html/test_html_to_ontology_parsing.py

@@ -356,12 +356,12 @@ def test_broken_cell_is_not_raising_error():
         """
     <div class="Page">
         <table class="Table">
-            <tbody class="TableBody">
-                <tr class="TableRow">
-                   <td class="TableCell" tablecell&quot;="">
+            <tbody>
+                <tr>
+                   <td tablecell&quot;="">
                     83.64 GiB
                    </td>
-                  <th class="TableCellHeader" rowspan="2">
+                  <th rowspan="2">
                     Fair Value
                  </th>
                </tr>
@@ -406,12 +406,12 @@ def test_table():
         """
     <div class="Page">
         <table class="Table">
-            <tbody class="TableBody">
-                <tr class="TableRow">
-                  <td class="TableCell">
+            <tbody>
+                <tr>
+                  <td>
                     Fair Value1
                   </td>
-                  <th class="TableCellHeader" rowspan="2">
+                  <th rowspan="2">
                     Fair Value2
                  </th>
                </tr>
@@ -467,24 +467,20 @@ def test_table_and_time():
         """
     <div class="Page">
         <table class="Table">
-            <thead class='TableHeader'>
-                <tr class="TableRow">
-                    <th class="TableCellHeader"  colspan="6">
+            <thead>
+                <tr>
+                    <th colspan="6">
                         Carrying Value
                     </th>
                 </tr>
             </thead>
-            <tbody class='TableBody'>
-                <tr class="TableRow">
-                    <td class="TableCell" colspan="5">
-                        <time class="CalendarDate">
+            <tbody>
+                <tr>
+                    <td colspan="5">
                             June 30, 2023
-                        </time>
                     </td>
-                <td class="TableCell">
-                    <span class="Currency">
+                <td>
                         $—
-                    </span>
                 </td>
                 </tr>
             </tbody>
@@ -594,3 +590,18 @@ def test_text_is_wrapped_inside_layout_element():
     parsed_ontology = indent_html(remove_all_ids(ontology.to_html()))
 
     assert parsed_ontology == expected_html
+
+
+def test_text_in_form_field_value():
+    # language=HTML
+    input_html = """
+    <div class="Page">
+    <input class="FormFieldValue" value="Random Input Value"/>
+    </div>
+    """
+    page = parse_html_to_ontology(input_html)
+
+    assert len(page.children) == 1
+    form_field_value = page.children[0]
+    assert form_field_value.text == ""
+    assert form_field_value.to_text() == "Random Input Value"

test_unstructured/partition/html/test_html_to_unstructured_and_back_parsing.py

@@ -314,24 +314,21 @@ def test_table():
     unstructured_elements, parsed_ontology = _parse_to_unstructured_elements_and_back_to_html(
         html_as_str
     )
-    expected_html = indent_html(html_as_str, html_parser="html.parser")
-    parsed_html = indent_html(parsed_ontology.to_html(), html_parser="html.parser")
 
-    assert expected_html == parsed_html
     expected_elements = _page_elements + [
         Table(
             text="Fair Value1 Fair Value2",
             detection_origin="vlm_partitioner",
             element_id="2",
             metadata=ElementMetadata(
                 text_as_html='<table class="Table" id="2"> '
-                '<tbody class="TableBody" id="3"> '
-                '<tr class="TableRow" id="4"> '
-                '<td class="TableCell" id="5">'
-                "Fair Value1 "
+                "<tbody> "
+                "<tr> "
+                "<td>"
+                "Fair Value1"
                 "</td>"
-                '<th class="TableCellHeader" rowspan="2" id="6">'
-                "Fair Value2 "
+                '<th rowspan="2">'
+                "Fair Value2"
                 "</th></tr></tbody></table>",
                 parent_id="1",
             ),

unstructured/__version__.py

@@ -1 +1 @@
-__version__ = "0.16.4-dev0"  # pragma: no cover
+__version__ = "0.16.4-dev1"  # pragma: no cover

unstructured/documents/ontology.py

@@ -20,6 +20,7 @@
 from enum import Enum
 from typing import List, Optional
 
+from bs4 import BeautifulSoup
 from pydantic import BaseModel, Field
 
 
@@ -75,32 +76,39 @@ def generate_unique_id() -> str:
 
     def to_html(self, add_children=True) -> str:
         additional_attrs = copy(self.additional_attributes)
-        if "class" in additional_attrs:
-            del additional_attrs["class"]
-
-        # TODO(Pluto) Add support for multiple classes
-        attrs = " ".join(
-            f'{key}="{value}"' if value else f"{key}" for key, value in additional_attrs.items()
-        )
+        additional_attrs.pop("class", None)
 
+        attr_str = self._construct_attribute_string(additional_attrs)
         class_attr = f'class="{self.css_class_name}"' if self.css_class_name else ""
-        attr_str = f"{class_attr} {attrs}".strip()
 
-        children_html = (
-            ("" if not self.children else "".join(child.to_html() for child in self.children))
-            if add_children
-            else ""
+        combined_attr_str = f"{class_attr} {attr_str}".strip()
+
+        children_html = self._generate_children_html(add_children)
+
+        result_html = self._generate_final_html(combined_attr_str, children_html)
+
+        return result_html
+
+    def to_text(self, add_children=True) -> str:
+        return " ".join(BeautifulSoup(self.to_html(add_children), "html.parser").stripped_strings)
+
+    def _construct_attribute_string(self, attributes: dict) -> str:
+        return " ".join(
+            f'{key}="{value}"' if value else f"{key}" for key, value in attributes.items()
         )
-        text = "" if not self.text else self.text
+
+    def _generate_children_html(self, add_children: bool) -> str:
+        if not add_children or not self.children:
+            return ""
+        return "".join(child.to_html() for child in self.children)
+
+    def _generate_final_html(self, attr_str: str, children_html: str) -> str:
+        text = self.text or ""
 
         if text or children_html:
-            # This is either one or another, never both
-            result_html = (
-                f"<{self.html_tag_name} {attr_str}>{text} {children_html}</{self.html_tag_name}>"
-            )
+            return f"<{self.html_tag_name} {attr_str}>{text} {children_html}</{self.html_tag_name}>"
         else:
-            result_html = f"<{self.html_tag_name} {attr_str} />"
-        return result_html
+            return f"<{self.html_tag_name} {attr_str} />"
 
     @property
     def id(self) -> str | None:
@@ -254,6 +262,18 @@ class Table(OntologyElement):
     elementType: ElementTypeEnum = Field(ElementTypeEnum.table, frozen=True)
     allowed_tags: List[str] = Field(["table"], frozen=True)
 
+    def to_html(self, add_children=True) -> str:
+        soup = BeautifulSoup(super().to_html(add_children), "html.parser")
+
+        for tag in soup.find_all(True):
+            if tag.name != "table":
+                tag.attrs.pop("class", None)
+                tag.attrs.pop("id", None)
+            if tag.name in ["td", "th"]:
+                tag.string = " ".join(tag.stripped_strings)
+
+        return str(soup)
+
 
 class TableBody(OntologyElement):
     description: str = Field("A body of the table", frozen=True)
@@ -430,6 +450,15 @@ class Form(OntologyElement):
     elementType: ElementTypeEnum = Field(ElementTypeEnum.form, frozen=True)
     allowed_tags: List[str] = Field(["form"], frozen=True)
 
+    def to_text(self, add_children=True) -> str:
+        texts = [self.text] if self.text else []
+
+        if add_children:
+            for child in self.children:
+                texts.append(child.to_text(add_children=True))
+
+        return " ".join(filter(None, texts)).strip()
+
 
 class FormField(OntologyElement):
     description: str = Field("A property value of a form", frozen=True)
@@ -442,6 +471,9 @@ class FormFieldValue(OntologyElement):
     elementType: ElementTypeEnum = Field(ElementTypeEnum.form, frozen=True)
     allowed_tags: List[str] = Field(["input"], frozen=True)
 
+    def to_text(self, add_children=True) -> str:
+        return super().to_text() + self.additional_attributes.get("value", "")
+
 
 class Checkbox(OntologyElement):
     description: str = Field("A small box that can be checked or unchecked", frozen=True)

unstructured/partition/html/transformations.py

@@ -96,10 +96,8 @@ def ontology_to_unstructured_elements(
         ]
         element_class = TYPE_TO_TEXT_ELEMENT_MAP[unstructured_element_class_name]
         html_code_of_ontology_element = ontology_element.to_html()
-        element_text = (
-            BeautifulSoup(html_code_of_ontology_element, "html.parser").get_text().strip()
-        )
-        # TODO value attribute from form input should be added to the text
+        element_text = ontology_element.to_text()
+
         unstructured_element = element_class(
             text=element_text,
             element_id=ontology_element.id,