Temporal Extraction

Documentation Index

Fetch the complete documentation index at: https://mintlify.com/intuit-ai-research/REMem/llms.txt

Use this file to discover all available pages before exploring further.

​

Overview

​

Class: TemporalExtraction

​

Constructor

from remem.information_extraction.temporal_extraction_openai import TemporalExtraction
from remem.llm.openai_gpt import CacheOpenAI

llm_model = CacheOpenAI(model="gpt-4")
extractor = TemporalExtraction(llm_model=llm_model, global_config=config)

​

Methods

​

batch_openie()

def batch_openie(
    self,
    chunks: Dict[str, ChunkInfo]
) -> Dict[str, TemporalRawOutput]

​

Parameters

​

Returns

​

Example Usage

from remem.information_extraction.temporal_extraction_openai import TemporalExtraction
from remem.llm.openai_gpt import CacheOpenAI

# Initialize the extractor
llm_model = CacheOpenAI(model="gpt-4")
config = type('Config', (), {
    'seed': 42,
    'temperature': 0.0
})()

extractor = TemporalExtraction(llm_model=llm_model, global_config=config)

# Prepare temporal chunks
chunks = {
    "event_1": {
        "metadata": {
            "timestamp": "2024-01-15",
            "content": "Marie Curie won the Nobel Prize in Physics in 1903 and the Nobel Prize in Chemistry in 1911. She worked at the University of Paris from 1906 to 1934."
        },
        "num_tokens": 70,
        "content": "Marie Curie won the Nobel Prize...",
        "chunk_order": [(0, 1)],
        "full_doc_ids": ["doc_789"]
    }
}

# Extract temporal facts
results = extractor.batch_openie(chunks)

# Access results
for chunk_id, output in results.items():
    print(f"\nTemporal facts in {chunk_id}:")
    for fact in output.facts:
        print(f"  Subject: {fact['subject']}")
        print(f"  Predicate: {fact['predicate']}")
        print(f"  Object: {fact['object']}")
        print(f"  Temporal: {fact['temporal_qualifier']}")
        print()

# Output:
#   Subject: Marie Curie
#   Predicate: won
#   Object: Nobel Prize in Physics
#   Temporal: 1903
#
#   Subject: Marie Curie
#   Predicate: won
#   Object: Nobel Prize in Chemistry
#   Temporal: 1911
#
#   Subject: Marie Curie
#   Predicate: worked at
#   Object: University of Paris
#   Temporal: from 1906 to 1934

​

Extraction Process

​

Temporal Fact Structure

1. Subject: The entity the fact is about
2. Predicate: The action, relationship, or state
3. Object: The target entity, value, or description
4. Temporal Qualifier: Time information in various formats:
   • Specific dates: “2024-01-15”, “January 2024”
   • Date ranges: “from 2020 to 2023”, “2020-2023”
   • Relative times: “last week”, “two years ago”
   • Durations: “for 3 months”, “since 2020”
   • Temporal relations: “before graduation”, “after the meeting”

​

Prompt Template

• Identify temporal expressions in the text
• Associate time information with facts
• Normalize temporal references when possible
• Handle various temporal formats and granularities

​

Passage Construction

• Structures temporal context appropriately
• Preserves timestamp information from metadata
• Formats content for temporal extraction

​

Temporal Qualifier Types

​

Absolute Time

• Specific dates: “2024-03-15”
• Specific times: “2024-03-15 14:30:00”
• Year-month: “March 2024”
• Year only: “2024”

​

Relative Time

• Past references: “yesterday”, “last week”, “2 days ago”
• Future references: “tomorrow”, “next month”, “in 3 weeks”
• Current references: “today”, “this week”, “currently”

​

Duration

• Explicit duration: “for 5 years”, “during 3 months”
• Open-ended: “since 2020”, “until now”
• Bounded: “from 2020 to 2023”, “between Jan and Mar”

​

Temporal Relations

• Sequential: “before X”, “after Y”, “during Z”
• Simultaneous: “while X”, “at the same time as Y”
• Frequency: “every Monday”, “twice a year”

​

Performance Considerations

• Parallel processing: Default max_workers=8 for ThreadPoolExecutor
• Single-stage extraction: Extracts complete temporal facts in one pass
• JSON mode: Structured output for reliable parsing
• Progress tracking: Real-time token usage and cache hit metrics
• Efficient threading: Balances concurrency with API rate limits

​

Error Handling

• Automatic JSON repair for truncated responses
• Graceful handling of malformed temporal expressions
• Returns empty fact lists on extraction failure
• Preserves error information in metadata
• Logs warnings for debugging
• Replaces null values with empty strings

​

Use Cases

​

Time-Aware Knowledge Graphs

# Build temporal knowledge representation
for fact in temporal_facts:
    kg.add_edge(
        fact['subject'],
        fact['object'],
        relation=fact['predicate'],
        timestamp=fact['temporal_qualifier']
    )

​

Event Timeline Construction

# Create chronological event sequences
events = sorted(
    temporal_facts,
    key=lambda f: parse_date(f['temporal_qualifier'])
)

​

Temporal Question Answering

# Query: "When did Marie Curie win the Nobel Prize?"
matching_facts = [
    f for f in temporal_facts
    if f['subject'] == 'Marie Curie'
    and 'Nobel Prize' in f['object']
]
print(matching_facts[0]['temporal_qualifier'])  # "1903"

​

Historical Analysis

# Extract facts within a time range
recent_facts = [
    f for f in temporal_facts
    if is_within_range(f['temporal_qualifier'], start='2020-01-01', end='2024-12-31')
]

​

Integration with Memory Systems

• Episodic memory: Link events to specific times
• Semantic memory: Track knowledge validity periods
• Autobiographical memory: Create personal timelines
• Context-aware retrieval: Filter by temporal constraints
• Memory decay modeling: Weight facts by recency