Claimify: Extracting high-quality claims from language model outputs
What this covers
Dasha Metropolitansky, Research Data Scientist, Microsoft Research Special Projects, introduces Claimify, a new method for extracting simple, verifiable claims from LLM outputs. Claim extraction is a key step in fact-checking LLM-generated content. Claimify outperforms prior techniques, ensuring that extracted claims are accurate, verifiable, and preserve critical context. Paper: https://www.microsoft.com/en-us/research/publication/towards-effective-extraction-and-evaluation-of-factual-claims/ Blog post: https://www.microsoft.com/en-us/research/blog/claimify-extracting-high-quality-claims-from-language-model-outputs/
Source description (no synthesized summary yet).
Claimify is a claim extraction system that decomposes long-form text into simple, verifiable factual statements to enable systematic evaluation of language model outputs, particularly for detecting hallucinations and assessing relevance and comprehensiveness.
- Claim extraction breaks down complex multi-paragraph answers into standalone factual statements that can be independently verified
- The system uses a three-stage pipeline (selection, disambiguation, decomposition) to handle ambiguity and extract high-quality claims
- This capability unlocks evaluation of long-form language model content by converting hard-to-assess paragraphs into measurable, verifiable units
This asset isn't compiled yet
You're seeing its claims, ranked. Compile it to build the argument threads, weight them, and check each claim against your library — the full view.
Language models in question-answering applications may generate hallucinations where they make up information rather than basing answers on source documents.
“So let's say that your question-answering application answers questions based on some source documents like news articles. You want to make sure that the language model is answering those questions based on the source documents, not just making things up.”
A claim is defined as a simple factual statement that can be verified as true or false.
“A claim, as I define it, is a simple factual statement that can be verified as true or false.”
One of the most popular use cases for language models is generating long-form content, and claim extraction can help with evaluation of such content.
“So, one of the most popular use cases of language models is generating long- form content. Unfortunately, it's really hard to evaluate the quality of that content. Claim extraction can help, and Claify is a really powerful tool for generating highquality claims.”
When extracting claims from the sentence 'some notable examples of technology executives include Satia Nadella and Bill Gates,' the word 'notable' should be removed because it cannot be verified as true or false, and separate independent claims should be created for each executive rather than combining them.
“One of them is that I got rid of this word notable because what does that even mean? that's not something that I can verify as true or false. The other is that I created separate claims, one for Bill Gates and one for Satia Nadella. This is because we want claims to be the simplest possible independent statements.”
Subjective qualifiers like 'notable' cannot be included in claims because they are not verifiable as true or false.
“One of them is that I got rid of this word notable because what does that even mean? that's not something that I can verify as true or false.”
Claims should be decomposed into the simplest possible independent statements rather than kept as compound claims.
“This is because we want claims to be the simplest possible independent statements.”
Hallucination detection in question-answering systems is difficult when answers contain multiple paragraphs of information, but can be made easier by distilling answers into simple standalone factual statements that can be checked independently.
“I'm working now on a system that does hallucination detection. So let's say that your questionans answering application answers questions based on some source documents like news articles. You want to make sure that the language model is answering those questions based on the source documents, not just making things up. But that's a really hard evaluation to perform when you have a paragraph or multiparagraph answer that has so much information in it. Now imagine if you could distill that into a simple set of standalone factual statements. It becomes much easier to then check those independently.”
Claimify processes text by breaking it down into sentences first, then performing claim extraction on each sentence independently, while including contextual windows of text around each sentence to ensure accurate interpretation.
“We don't try to do claim extraction on the whole text at once. We actually break it down into sentences and we do the claim extraction on each sentence independently. Now to ensure that those sentences are interpreted accurately, we include some context which is basically a window of text around the sentence.”
When Claimify encounters the sentence about contaminated water and improved water management, it determines that the context does not clearly support either interpretation and flags the sentence as unable to be disambiguated, declining to proceed to the decomposition stage.
“In other words, this may or may not be a verifiable claim. Claimifi decided that the context did not clearly support either interpretation. So it flagged the sentences cannot be disambiguated and did not proceed to the decomposition stage.”
When a language model generates the sentence 'the UN found that the resulting contaminated water caused many residents to fall ill, highlighting the need for improved water management,' a baseline system may miss the ambiguity about whether the author is stating a verifiable fact (that the UN found the need for improved water management) or adding their own interpretation.
“The baseline prompt ignored the phrase highlighting the need for improved water management. It only extracted claims from the first part of the sentence. However, Claify reasoned, quote, the sentence could be interpreted as the UN found that the contaminated water caused illness and also highlighted the need for improved water management. Or it could be interpreted as the UN only found the contamination and illness and the author is adding the interpretation about the need for improved water management.”
Claim extraction enables evaluation of answer relevance by allowing assessment of individual claims within an answer, where some claims may be relevant to a question while others are not, enabling aggregation into a composite relevance measure.
“You want to know how relevant the answers are to the question that was asked. If this answer contains 20 or 30 distinct points, it's so hard to say how relevant the entire answer is. Maybe some points are relevant and others aren't. But if you can take these individual factual claims and say this one is relevant, this one isn't. You can easily aggregate that into one composite measure.”
Claimify's ability to determine whether ambiguity in a sentence can be resolved using context is a particularly unique and powerful aspect of the disambiguation stage.
“This is one aspect of claimify that is really unique and powerful, especially the ability to determine whether or not this ambiguity can be resolved.”
In the selection stage of Claimify, sentences that do not contain verifiable claims are filtered out, such as opinion statements like 'companies should embrace AI.'
“Selection means we're filtering out sentences that do not contain any verifiable claims. So for example, if I give you the sentence, companies should embrace AI, that's not a factual claim. It's an opinion. So we would filter it out.”
Claim extraction matters because it unlocks the ability to evaluate long-form content generated by language models.
“So to summarize, why does claim extraction matter? Because it unlocks the ability to evaluate long- form content generated by language models.”
Claimify breaks down claim extraction into three distinct stages: selection (filtering sentences without verifiable claims), disambiguation (detecting and resolving ambiguity), and decomposition (breaking sentences into simple factual statements).
“Number two is we don't treat claim extraction as one monolithic task. We break it down into three parts. Selection, disambiguation, and decomposition.”
When an answer contains 20 or 30 distinct points, it is difficult to assess how relevant the entire answer is to the question because individual points may vary in relevance.
“If this answer contains 20 or 30 distinct points, it's so hard to say how relevant the entire answer is. Maybe some points are relevant and others aren't.”
Claimify does not perform claim extraction on the entire text at once; instead, it breaks the text down into sentences and performs claim extraction on each sentence independently.
“We don't try to do claim extraction on the whole text at once. We actually break it down into sentences and we do the claim extraction on each sentence independently.”
Claimify includes context—a window of text around each sentence—to ensure accurate interpretation of sentences during claim extraction.
“Now to ensure that those sentences are interpreted accurately, we include some context which is basically a window of text around the sentence.”
Claimify breaks down claim extraction into three parts: selection, disambiguation, and decomposition.
“Number two is we don't treat claim extraction as one monolithic task. We break it down into three parts. Selection, disambiguation, and decomposition.”
Evaluating hallucination detection is difficult when answers are paragraph or multi-paragraph length because it's hard to systematically verify all the information.
“But that's a really hard evaluation to perform when you have a paragraph or multi-paragraph answer that has so much information in it.”
Claim extraction enables evaluation beyond hallucination detection, including assessment of answer relevance to questions.
“But this is not just about hallucination detection. You can do other sorts of evaluations as well.”
Individual factual claims can be assessed independently for relevance and then aggregated into a composite relevance measure for the entire answer.
“But if you can take these individual factual claims and say this one is relevant, this one isn't. You can easily aggregate that into one composite measure.”
Disambiguation in the Claimify pipeline is the process of detecting ambiguity in sentences and determining whether ambiguity can be resolved using context or must be flagged as unresolvable.
“Secondly, we have disambiguation. This is basically detecting whether there's ambiguity and then deciding if there's ambiguity, can it be resolved using the context or flag that it can't be resolved.”
Distilling multi-paragraph answers into simple standalone factual statements makes it easier to check hallucinations independently.
“Now imagine if you could distill that into a simple set of standalone factual statements. It becomes much easier to then check those independently.”
Ambiguity in claim extraction refers to situations where multiple plausible interpretations of a sentence exist, and depending on which interpretation is chosen, a very different set of claims will result.
“Secondly, we have disambiguation. This is basically detecting whether there's ambiguity and then deciding if there's ambiguity, can it be resolved using the context or flag that it can't be resolved. Ambiguity here just means there are multiple plausible interpretations. And depending on which interpretation you pick, you're going to get a very different set of claims.”
When processing the sentence 'countries like Afghanistan and Sudan have experienced similar challenges to those of Libya,' a baseline system failed to specify what 'those' refers to, but Claimify made specific claims about the actual issues (public health crisis, flooding, and contaminated water) mentioned in the context.
“Consider the sentence countries like Afghanistan and Sudan have experienced similar challenges to those of Libya where the baseline claims never specified what those refers to. While the context discusses public health crisis, flooding and contaminated water. So Claifi made specific claims about these issues.”
When processing the sentence 'Nigeria is striving to become self-sufficient in wheat production, but is hindered by climate change and violence,' Claimify correctly captured that it is Nigeria's efforts to achieve self-sufficiency that are being hindered, whereas the baseline only stated that wheat production itself is hindered.
“In the sentence, Nigeria is striving to become self-sufficient in wheat production, but is hindered by climate change and violence. The baseline had claims like Nigeria's wheat production is hindered by climate change and violence. Claimifi captured that it's Nigeria's efforts to become self-sufficient in wheat production that are being hindered.”
When processing the sentence 'Argentina's currency value has plunged,' Claimify correctly specifies that inflation is the cause of the currency depreciation, whereas the baseline only stated that the currency value has plunged without identifying the causal mechanism.
“Also, the baseline just said Argentina's currency value has plunged. Claimifi correctly specified that inflation has depreciated the currency.”
The number of claims in an answer can be used as a proxy for how comprehensive it is.
“Our team is also using the number of claims in the answer as a proxy for how comprehensive it is.”
The ability to determine whether ambiguity in text can be resolved using context is a unique and powerful aspect of Claimify.
“This is one aspect of claimify that is really unique and powerful, especially the ability to determine whether or not this ambiguity can be resolved.”
Claimify is a system that takes text of any length and decomposes it into high-quality claims.
“So, taking a step back, claimify basically takes in a text of any length. Usually, these texts are much longer than the example I just gave you, which was just a sentence, and it'll take that text and decompose it into these highquality claims.”
Claim extraction is important because it unlocks the ability to evaluate long-form content generated by language models.
“To summarize, why does claim extraction matter? Because it unlocks the ability to evaluate long-form content generated by language models.”
Ambiguity in text means there are multiple plausible interpretations, and depending on which interpretation is chosen, different sets of claims will be extracted.
“Ambiguity here just means there are multiple plausible interpretations. And depending on which interpretation you pick, you're going to get a very different set of claims.”
The number of claims in an answer can be used as a proxy metric for how comprehensive an answer is.
“Our team is also using the number of claims in the answer as a proxy for how comprehensive it is.”
In comparison to a baseline system, Claimify correctly extracted claims about Argentina's inflation causing economic hardship and predictions of inflation rates greater than 300%, while the baseline missed these claims.
“Recall the sentences about Argentina's inflation where the baseline missed the claims about economic hardship and the prediction of rates greater than 300%. Claimifi did not miss these claims.”
By processing the entire chatbot answer at once, a baseline claim extraction approach misses claims about economic hardship and predictions of inflation rates greater than 300% in Argentina.
“Recall the sentences about Argentina's inflation where the baseline missed the claims about economic hardship and the prediction of rates greater than 300%.”
Claimify correctly extracted claims about economic hardship and predictions of inflation rates greater than 300% in Argentina, whereas the baseline missed these claims.
“Claimifi did not miss these claims.”
The baseline approach said 'Argentina's currency value has plunged' but did not specify the mechanism, while Claimify correctly specified that inflation has depreciated the currency.
“Also, the baseline just said Argentina's currency value has plunged. Claimifi correctly specified that inflation has depreciated the currency.”
For the sentence 'Nigeria is striving to become self-sufficient in wheat production, but is hindered by climate change and violence', Claimify correctly captured that it is Nigeria's efforts to become self-sufficient that are being hindered, rather than just the wheat production itself.
“In the sentence, Nigeria is striving to become self-sufficient in wheat production, but is hindered by climate change and violence. The baseline had claims like Nigeria's wheat production is hindered by climate change and violence. Claimifi captured that it's Nigeria's efforts to become self-sufficient in wheat production that are being hindered.”
Claim extraction can help evaluate the quality of long-form language model content, and Claimify is a powerful tool for generating high-quality claims.
“Claim extraction can help, and Claify is a really powerful tool for generating highquality claims.”
The baseline claim extraction approach treated the entire chatbot answer as a monolithic unit and did not use context-aware sentence-level processing.
“The baseline prompt ignored the phrase highlighting the need for improved water management. It only extracted claims from the first part of the sentence.”
Claimify identified an ambiguity in the sentence 'the UN found that the resulting contaminated water caused many residents to fall ill, highlighting the need for improved water management' by reasoning that it could be interpreted either as the UN finding both the contamination/illness AND the need for improved water management, or as the UN only finding contamination/illness while the author added interpretation about improved water management.
“Claify reasoned, quote, the sentence could be interpreted as the UN found that the contaminated water caused illness and also highlighted the need for improved water management. Or it could be interpreted as the UN only found the contamination and illness and the author is adding the interpretation about the need for improved water management.”
Because Claimify could not determine from the context which interpretation was correct regarding whether the UN found the need for improved water management or the author was adding that interpretation, it flagged the sentence as unable to be disambiguated and did not proceed to the decomposition stage.
“In other words, this may or may not be a verifiable claim. Claimifi decided that the context did not clearly support either interpretation. So it flagged the sentences cannot be disambiguated and did not proceed to the decomposition stage.”
For the sentence 'countries like Afghanistan and Sudan have experienced similar challenges to those of Libya', the baseline never specified what 'those' refers to, while Claimify made specific claims about the issues (public health crisis, flooding, contaminated water) based on context.
“Consider the sentence countries like Afghanistan and Sudan have experienced similar challenges to those of Libya where the baseline claims never specified what those refers to. While the context discusses public health crisis, flooding and contaminated water. So Claifi made specific claims about these issues.”
In a chatbot answer about challenges in emerging markets, assessing quality is difficult because the answer is packed with information and there is no gold standard to compare against.
“Assessing the quality of the answer is really hard. It's packed with information and there's no gold standard to compare against.”
Extraction is the process of breaking down a text into claims.
“And extraction is the process of breaking down a text into claims.”
Claimify takes text of any length and decomposes it into high-quality claims.
“Claimify basically takes in a text of any length. Usually, these texts are much longer than the example I just gave you, which was just a sentence, and it'll take that text and decompose it into these high-quality claims.”
One of the most popular use cases of language models is generating long-form content.
“So, one of the most popular use cases of language models is generating long-form content.”
The decomposition stage of Claimify takes a disambiguated sentence and breaks it down into simple, standalone factual statements.
“And then the last stage is decomposition, which takes the disambiguated sentence and breaks it down into these simple standalone factual statements.”
A claim is defined as a simple factual statement that can be verified as true or false.
“A claim, as I define it, is a simple factual statement that can be verified as true or false.”
Extraction is the process of breaking down a text into claims.
“And extraction is the process of breaking down a text into claims.”
It is difficult to evaluate the quality of long-form content generated by language models.
“Unfortunately, it's really hard to evaluate the quality of that content.”
Selection in the Claimify pipeline means filtering out sentences that do not contain any verifiable claims.
“Selection means we're filtering out sentences that do not contain any verifiable claims.”
An example of a sentence that would be filtered out in the selection stage is 'companies should embrace AI' because it is an opinion, not a factual claim.
“So for example, if I give you the sentence, companies should embrace AI, that's not a factual claim. It's an opinion. So we would filter it out.”
Decomposition in the Claimify pipeline takes the disambiguated sentence and breaks it down into simple standalone factual statements.
“And then the last stage is decomposition, which takes the disambiguated sentence and breaks it down into these simple standalone factual statements.”