Can GPT Redefine Medical Understanding?
Evaluating GPT on Biomedical Machine Reading Comprehension

The prompt template used for this technique is shown in Figure 1. The Basic prompting approach asks GPT to answer the query in the simplest way possible. The profession placeholder specifies the role that GPT has to take in order to answer the asked question. Based on the source of the dataset, this placeholder takes the value of biologist in the case of ProcessBank, biomedical researcher in the case of BioMRC, consumer healthcare expert in case of MASH-QA and medical expert in the case of CliCR dataset. Again, based on the source of the dataset, the placeholder context_type takes the value of paragraph for ProcessBank, abstract of the paper for BioMRC, healthcare article for MASH-QA and clinical case report for CliCR. The placeholder query_type takes the value of query for ProcessBank, title containing the missing entity for BioMRC, query for MASH-QA and query containing the missing entity for CliCR. The query_text placeholder contains the actual text of the query and similarly context_text contains the actual text of the context. The options placeholder is present only for ProcessBank and BioMRC datasets and contains the choices to select from while answering the asked query.

The rationale behind using the CoT technique is that there may be multiple smaller questions that need to be answered first in order to conclude the answer of the final asked question. For example, one of the questions asked to GPT is Has there been at least 6 weeks of provider-directed conservative treatment?. This question can easily be divided into 3 smaller questions Has there been any conservative treatment?, Was the treatment provider-directed? and What was the duration of conservative treatment?. The prompt template used for this technique is exactly same to that of Figure 1 with an additional line instructing the model to Think step by step.

Inspired by Yasunaga et al. (2023b), we design our own analogical reasoning strategy by tweaking the prompt to fit our problem statement. We do this because, unlike the general domain, GPT would not be able to recall specific dataset-level knowledge, as we are not sure if it was ever trained on the datasets being used in our study. Rather, we frame the prompt so that GPT does not need to rely on a lot on global knowledge. To that end, instead of asking GPT to generate any kind of relevant QA pairs based on its global knowledge, we ask GPT to generate QA pairs from the given context and then answer the initial question. There is one hyperparameter for this technique which is the number of QA pairs to generate.

Most of the work on RAG talks about data retrieval based on an accepted relevancy score and then using LLM prompts to answer the given query. The data retrieval is done by storing the embeddings from some encoder of the entire corpus (a datapoint’s context in our case) in a vector database index and then retrieving the most matching data points (text extracts or sections from a datapoint’s context) for a given query. The key idea behind using RAG is that it helps in saving a lot of computational cost and improves the LLM’s performance as now it has to look in a smaller knowledge space to answer the asked question. In our proposed novel prompting technique Implicit RAG, we completely ignore the overhead involved in getting the embeddings of the entire corpus and storing them in a vector database. Instead, we ask the LLM itself to find the most relevant text extracts or sections in the given context which may help in answering the asked question, and then later use these extracted sections to conclude the answer to the original question. The general working of our proposed prompting technique is shown in Figure 2. There are two hyper-parameters for this technique. First is, the number of sections to extract, and the next one is the number of words in each section or text extract. The prompt template used for this technique is shown in Figure 3. The hyper-parameter values for the number of sections and number of words in each section is provided in the placeholders number_of_sections and lower_limit_length and upper_limit_length.

The results for ProcessBank are shown in Table 2. We ran all 4 prompting strategies on the entire test set of 150 datapoints in a zero-shot setup. Every single prompting method outperforms the previously proposed methods, thus giving us a new SoTA on this dataset. Among the different prompting strategies, Implicit RAG gets the best results. The important observations are:

1. The only 4 to 5 datapoints that GPT got wrong either are very confusing for even humans to answer or had some typo or extra punctuation in ground-truth answers which GPT was not able to mimic during its generation.

2. It is observed that all the GPT prompting strategies work more or less the same if the question can be answered from a small span in the provided context. The reason why Implicit RAG is able to outperform other techniques is because this dataset includes around 30% of temporal and true-false type questions which require extensive analysis of the entire context and that the answer can be spread in different segments of the context. Therefore, reducing the knowledge space by extracting relevant sections to answer the asked question helps in improving the performance.

The results for BioMRC are listed in Table 3. Due to cost-related reasons, we first compare different prompting methods by running them on a randomly selected 15% (1000 datapoints) subset of the test set and then choosing the best prompting technique to run on the entire test set. All these experiments are done in a zero-shot setting. Amongst the different prompting techniques, Basic prompting gets the best results and Implicit RAG ranks second. The important observations are:

1. Even though BioMRC is a cleaner version of BioREAD, there are still elements of lack of structure in the dataset. For example, there is no 1-1 mapping between entity IDs and entities. So, this means the same entity can be mapped to multiple entity IDs and vice versa which causes a lot of confusion when quantifying performance. The authors of BioMRC claim that for any query, the abstract or the context contains all the candidate options including the correct answer but this is not always true leading to more confusion during evaluation.

2. Quite a few times GPT is able to generate an acronym answer instead of its corresponding full form. Ideally, both acronyms and their full forms must be considered as correct answers.

3. There are instances where GPT being a generative model is able to produce semantically similar answers but still they are marked wrong as they do not exactly match with the correct answer. An embedding based metric can be helpful here.

4. There are a lot of entities which are semantically and syntactically the same but still belong to different ontologies and thus have different entity IDs. For example, in one case, GPT generates the answer amino acids where the correct answer is amino acid but since both of these entities have different IDs, this answer had to be marked wrong.

Another important aspect to note here apart from the lack of structure in the dataset is the overall system design of supervised models which are being compared to GPT when talking about SoTA. Supervised models use 70%-80% of the available data as their training set which allows their parameters to get a good idea about the nuances of the dataset whereas in case of GPT, all our experiments are being conducted in a zero-shot setting. Also since GPT is a generative model, the chances of GPT generating an answer not present in the candidate answer list despite the final answer being semantically and syntactically the same is really high. But a supervised model is never going to face this problem as it makes its prediction based on the confidence score for each candidate answer and thus the final answer is always going to be present in the candidate answer list.

The results for MASH-QA are shown in Table 4. We start by conducting a comparison between different prompting strategies by evaluating them on a randomly selected 15% (600 datapoints) subset of the test set. These experiments are undertaken in a zero-shot setting. As we can see in Table 4, Basic prompting performs the best while Implicit RAG ranks second. The important points to discuss here are:

1. The answers in QA pairs of MASH-QA are very subjective. The authors of this dataset have not specified any structured process that was followed by the healthcare experts when trying to answer the questions asked on a website from where this dataset was sourced in the first place. An in-depth analysis shows that even though GPT is able to extract better answers a lot of times, since it does not match with the ground truth answers, the evaluation metrics do not reflect it’s true capabilities.

2. There is a correlation observed between increase in the number of sections and decreasing performance of Implicit RAG. The reason is that the answers in this dataset are long span and thus with increasing number of sections, there is a loss of contextual continuity as the ground truth answers can get split across multiple sections. This ends up confusing the LLM making it difficult to choose the right set of sentences from different sections. Hence, it performs the best when asked to extract just one section from the context. But because MASH-QA contains answers which can be present in disjoint spans of the context, extracting just one section is not able to make Implicit RAG the best performing prompting method.

3. One question which may arise is whether extracting one section or having the hyper-parameter number of sections set to 1 for Implicit RAG makes it same as that of Basic prompting. Implicit RAG and Basic prompting become the same only when we are not only extracting just one section from the context but also when the hyper-parameter number of words for Implicit RAG is set equal to the length of the entire given context. But for MASH-QA, the hyper-parameter number of words is set to 300 with the lower limit being 0 and upper limit being 300 and hence they are different.

Again, we need to reiterate that GPT’s performance in case of MASH-QA is being compared to supervised models which use 70%-80% of the total data as their training set allowing their parameters to capture granular details better than a generic LLM like GPT in a zero-shot setting.

The results for CliCR can be seen in Table 5. Again, due to cost-related reasons, we first compare different prompting techniques by running them on randomly selected 15% (1100 datapoints) subset of the test set and then choosing the best prompting method to run on the entire test set. All these experiments are done in a zero-shot setting. Amongst the different prompting strategies, Implicit RAG and AR get the best results in terms of F1 metric. AR minutely performs better than Implicit RAG when compared in terms of the Exact Match (EM) metric. However, EM is a very harsh metric for a generative model as there can be so many possible variations of semantically similar output which are not wrong. Since AR is computationally faster with respect to Implicit RAG, we ran AR on the entire dataset. All the prompting methods outperform the previously proposed methods. Not only does GPT surpass the performance of previous models, but it also comes close to Human Expert performance while beating Human Novice results. The important observations are:

1. The authors of CliCR mention that for the training of supervised models, only those instances are used for which at least one ground-truth answer from the set of ground-truth answers occurs in the clinical case report or the context. But for the evaluation part, for both validation and test sets even those datapoints are included where there is no intersection between ground-truth set and the entities mentioned in the context. This favors supervised learning settings as supervised models have a separate training and development set which can allow the parameters of the model to learn such cues. GPT is still able to perform better possibly because the global knowledge embedded in its parameters gives it enough evidence to perform well.

2. The authors of CliCR compare various skills in their work between the previous SoTA (GPT is the new SoTA) model GA-NoEnt and Human Expert and show that there still exists a huge gap between them. Since GPT is able to achieve almost Human Expert level performance, we can expect it to show similar capabilities in other MRC tasks.

3. There are multiple reasons why Implicit RAG performs well on this dataset. First, the mean length of context in this dataset is 1461 words which indicates that with increasing size of context, the chances of analysis of different sections of the context simultaneously to answer a question is high and that is the core idea behind Implicit RAG. Next, the authors of CliCR list out that 70% of the queries in this dataset require the bridging skill, 40% require the skill of tracking and around 25% demand the spatiotemporal skill. All these three skills indicate that answering queries in this dataset require deriving cues from different segments of the context and that is what we propose as the key rationale behind Implicit RAG.

Out of the four datasets that we use in this study, Implicit RAG is able to achieve the best results for two of them when compared with other prompting techniques. It ranks at the second place for the other two datasets. One of the questions which may arise is whether Implicit RAG can be applicable to contexts which cannot fit in LLM’s 32k token limit. Implicit RAG will especially perform better than other prompting techniques in cases where the context size is more than 32k. In such cases, we can chunk the context and make multiple calls to Implicit RAG to retrieve relevant sections given the query. Once all the relevant sections have been retrieved, the last call to Implicit RAG can use these sections to arrive to an answer. But all other prompting techniques require analysis of the entire context (greater than 32k in this case) at the same time to arrive to an answer. We further do a qualitative analysis on 50 randomly picked datapoints for all four datasets. We check how many times the extracted sections are relevant to the question or not. Even if 1 out of all the extracted sections are relevant, we consider that to be a valid retrieval irrespective of whether the final answer was correct or incorrect. The results are shown in Table 6. As we can see, Implicit RAG is able to extract relevant sections in most cases.