Science

Many information retrieval tasks require viewing documents in some manner, whether this is to view information in context or to provide annotations for some downstream task (e.g., evaluation or system training). Building a high-quality document viewer often exceeds the resources of many researchers and so, in this paper, we describe the design and architecture of our new open-source document viewer, Spectator. In particular, we provide a look into the algorithmic details of how Spectator accomplishes tasks like mapping annotations back to the canonical document. Moreover, we provide a sampling of the use cases that we envision for Spectator, potential future additions depending on community need and support, and highlight situations where Spectator may not be a good fit. Furthermore, we provide a brief description of the sample application that we bundle with Spectator to demonstrate how one might use it within the context of a larger system.

In Mergers and Acquisition due diligence, lawyers are tasked with analyzing a collection of contracts and determine the level of risk that comes from a merger or acquisition. This process has historically been manual and resulted in only a small fraction of the collection being examined. This paper reports on the user-focused redesign of our document viewer that is used by clients to review documents and train machine learning algorithms to find pertinent information from these contracts.

While document signatures are a well established tool in IR, they have primarily been investigated in the context of web documents. Legal due diligence documents, by their nature, have more similar structure and language than we may expect out of standard web collections. Moreover, many due diligence systems strive to facilitate real-time interactions and so time from document ingestion to availability should be minimal. Such constraints further limit the possible solution space when identifying near duplicate documents. We present an examination of the tradeoffs that document signature methods face in the due diligence domain. In particular, we quantify the trade-off between signature length, time to compute, number of hash collisions, and number of nearest neighbours for a 90,000 document due diligence corpus.

We are concerned with investigating the apparent effective-ness of Radford et al.’s “Sentiment Neuron,” which they claim encapsulates sufficient knowledge to accurately predict sentiment in reviews. In our analysis of the Sentiment Neuron, we find that the removal of the neuron only marginally affects a classifier’s ability to detect and label sentiment and may even improve performance. Moreover, the effectiveness of the Sentiment Neuron can be surpassed by simply using 100 random neurons as features to the same classifier. Using adversarial examples, we show that the generated representation containing the Sentiment Neuron (i.e., the final hidden cell state in a LSTM) is particularly sensitive to the end of a processed sequence. Accordingly, we find that caution needs to be applied when interpreting neuron-based feature representations and potential flaws should be addressed for real-world applicability.

In due diligence, lawyers are tasked with reviewing a large set of legal documents to identify documents and portions thereof that may be problematic for a merger or acquisition. In an effort to aid users to review more efficiently, we sought to determine how document-level clustering may help users of a due diligence system during their workflow. Following an iterative design methodology, we conducted several user studies with different versions of a document-level clustering feature consisting of three distinct phases and 27 users. We found that the interface should adapt to a user’s understanding of what “similar documents” means so that trust can be established in the feature. Furthermore, the ability to negotiate with the underlying algorithm is facilitated by the establishment of trust. Finally, while the usage of this feature may be influenced by a user’s role, it remains primarily a project management tool.

As professional users interact with more AI-enabled tools, it has become increasingly important to understand how their work and behaviour are affected by such tools. In this paper, we present the insights that we have gleaned from a qualitative user study conducted with nine of our software’s users who are all legal professionals. We find that as our participants become more accustomed to the system they begin to subtly alter their behaviours and interactions with the system. Using their shared experiences, we distill these into insights that may inform the design of similar systems.

In the TREC Total Recall Track (2015-2016), participating teams could employ either fully automatic or human-assisted (“semi-automatic”) methods to select documents for relevance assessment by a simulated human reviewer. According to the TREC 2016 evaluation, the fully automatic baseline method achieved a recall-precision breakeven (“R-precision”) score of 0.71, while the two semi-automatic efforts achieved scores of 0.67 and 0.51. In this work, we investigate the extent to which the observed effectiveness of the different methods may be confounded by chance, by inconsistent adherence to the Track guidelines, by selection bias in the evaluation method, or by discordant relevance assessments. We find no evidence that any of these factors could yield relative effectiveness scores inconsistent with the official TREC 2016 ranking.

The problem of answering multiple choice questions, based on the content of documents has been studied extensively in the machine learning literature. We pose the due diligence problem, where lawyers study legal contracts and assess the risk in potential mergers and acquisitions, as a multiple choice question answering problem, based on the text of the contract. Existing frameworks for question answering are not suitable for this task, due to the inherent scarcity and imbalance in the legal contract data available for training. We propose a question answering system which first identifies the excerpt in the contract which potentially contains the answer to a given question, and then builds a multi-class classifier to choose the answer to the question, based on the content of this excerpt. Unlike existing question answering systems, the proposed system explicitly handles the imbalance in the data, by generating synthetic instances of the minority answer categories, using the Synthetic Minority Oversampling Technique. This ensures that the number of instances in all the classes are roughly equal to each other, thus leading to more accurate and reliable classification. We demonstrate that the proposed question answering system outperforms the existing systems with minimal amount of training data.

We present and formalize the due diligence problem, where lawyers extract data from legal documents to assess risk in a potential merger or acquisition, as an information retrieval task. Furthermore, we describe the creation and annotation of a document collection for the due diligence problem that will foster research in this area. This dataset comprises 50 topics over 4,412 documents and ~15 million sentences and is a subset of our own internal training data.